Main

We sought to characterize the potential regulatory contributions of different TFs to combinatorial enhancer control, that is, the regulatory functions of the TF proteins following DNA binding, regardless of their specific roles in vivo. We reasoned that such contributions could be best assessed using ectopic tethering assays in the context of DNA sequences that closely resemble active enhancers, ideally only lacking the input of a single TF. In particular, such a setup may allow the assessment of obligate combinatorial factors whose regulatory activities depend on partners and which would otherwise appear non-functional. We therefore developed enhancer complementation assays based on activator bypass experiments4 used to test candidate TF or cofactor function in transcription control and to dissect promoters5,6,7,8,9,10.

We mutated TF-binding-motif sequences within active enhancers to ‘upstream activating sequence’ (UAS) motifs for the GAL4 DNA-binding domain (GAL4-DBD), recruited 474 Drosophila TFs11 via GAL4-DBD fusion proteins to the positions of the mutated motifs (enhancer context), and measured enhancer activities by luciferase assays in S2 cells, normalizing to GFP recruitment (Fig. 1a). Since expression and recruitment is standardized, the factors’ regulatory functions can be assessed in a highly controllable manner independently of the factors’ endogenous expression and DNA binding. Overall, the assays were highly reproducible: 75% of all data points had standard deviations (s.d.) <10%, and 95% of the s.d. were <18% across four biological replicates.

Figure 1: Enhancer complementation assays for 474 TFs.
figure 1

a, Schematic overview. b, Activity of an enhancer and motif-mutant variants (data from ref. 3). a.u., arbitrary units. c, Enhancer complementation assays for CGCG- and GATA-contexts (normalized luciferase values for 474 TFs; red: activation, blue: repression). d, Preferential activation of CGCG-versus GATA-context (≥1.5-fold, FDR-corrected P < 0.05). Black boxes, TFs inactive on their own; colour, TFs tested in e. e, Validation of context-dependent TFs (details see text). Luciferase activities (firefly/Renilla) with (colours) or without (grey and black (Ets21C is expressed in S2 cells)) co-transfection of the respective untagged TF. Error bars show standard deviations (n = 4, biological replicates).

PowerPoint slide

Full size image

We started with an enhancer that was highly active in Drosophila S2 cells and for which mutations of CGCG- or GATA-type motifs strongly reduced its activity3 (Fig. 1b). We replaced either the CGCG- or the GATA-motifs with UAS motifs (Fig. 1a) and assessed which TFs restored enhancer function or repressed the remaining basal activities. Of the 474 TFs, 100 were activating (≥1.5-fold compared to GFP; P < 0.05 false discovery rate (FDR)-corrected for 474 tests) in the CGCG- and 84 in the GATA-context (Fig. 1c), including TFs that recognize the CGCG- and GATA-motifs, respectively (Extended Data Table 1). This compares to 77 TFs that activated on their own (that is, when recruited to UAS motifs outside an enhancer context), suggesting that TF function might be context-dependent. Indeed, 46 TFs activated the CGCG context at least 1.5-fold (P < 0.05) more strongly than the GATA context, even though both contexts were derived from the same enhancer (Fig. 1d and Extended Data Fig. 1). To test if native untagged TFs recapitulate these results, we chose Ets at 21C (Ets21C), Deformed (Dfd), and Hairy/E(spl)-related with YRPW motif (Hey) that preferentially activated the CGCG context to different extents (Fig. 1d). We replaced the CGCG- and GATA-motifs with binding sites for these TFs and expressed the untagged TFs (Fig. 1e). This activated the mutant enhancers in a manner consistent with the results from GAL4-DBD-mediated recruitment: Ets21C and Dfd activated only the CGCG context, while Hey activated both (CGCG 1.3-fold more highly), confirming the similarity and context-dependency of these TFs’ regulatory functions.

Intrigued by the context-dependency of some TFs even within a single enhancer, we decided to include more diverse regulatory contexts (Extended Data Fig. 1). We created 19 motif-mutant enhancer contexts for different types of TF motifs and different enhancers with broad, cell-type-specific3, or hormone-inducible12 activities. We also added five contexts consisting of UAS sites and core promoters specific towards developmental or housekeeping enhancers, respectively13.

Nearly half of all TFs (42%) were activating and most (93%) of the remaining 276 TFs were repressing in at least one of the 24 contexts (≥1.5-fold; P < 0.05 FDR-corrected for 24 × 474 tests), suggesting that most TF-fusion proteins were functional. Many TFs had similar regulatory effects across the 24 contexts, suggesting that they might be functionally equivalent. We grouped all TFs into 15 clusters using unsupervised spectral clustering (Fig. 2a and Supplementary Table 1) and confirmed that these clusters are robust to bootstrapping and reproducible when using independent biological replicates (Extended Data Fig. 2). This revealed clusters of diverse regulatory functions (Fig. 2b and Extended Data Fig. 3), including cluster 8 with TFs that activated in most contexts (global activators) such as Antennapedia, Sox14 and Sox15, Clock (Clk), and Zelda, and clusters 3, 5 and 7 with global repressors (for example, Snail, Runt, Engrailed and Kruppel). These TFs seemed to dominate or override other regulatory cues, consistent with their ability to function in isolation. TFs of other clusters were only weakly active in the contexts tested, and, notably, the TFs of some clusters were context-dependent. For example, cluster 10 TFs preferentially activated the housekeeping core promoter and might constitute factors of a distinct transcriptional program13, including Myb-interacting protein 120 (Mip120) and CG6813, the cluster’s strongest activator (21-fold; P = 4.4 × 10−4). In contrast, cluster 1 TFs preferentially activated hormone-receptor contexts, that is, when recruited to ecdysone receptor (EcR)-binding sites in enhancers inducible by the insect steroid hormone ecdysone. Examples are Twist, Reversed polarity, Pointed, and other developmental TFs.

Figure 2: TFs have diverse regulatory functions.
figure 2

a, 15 TF clusters (see text for highlighted clusters). b, Normalized luciferase values across 24 contexts (selected from Extended Data Fig. 3). c, Validation of hormone-context-preferential TFs (luciferase activities (firefly/Renilla) for re-engineered enhancers (re-eng.; purple) and controls (grey; see schematic)). Error bars show standard deviations (n = 4, biological replicates). **P < 1 × 10−2; ***P < 1 × 10−3. d, Enrichments for TFs that activate or repress the 4×UAS-dCP or 4×UAS-hkCP contexts (‘Functional’), protein-sequence features and GO-categories; significant (P < 0.01; empirical FDR = 0.1) enrichments, red; and depletions, blue; others, white; see Extended Data Fig. 4 and Supplementary Table 2 for details and all data. e, Pairwise distances between activity profiles in Kc167, BG3 and OSC cells support functional TF clusters (all empirical P < 10−6, indicated with single asterisks).

PowerPoint slide

Full size image

Intrigued by TFs that preferentially activated hormone-receptor contexts, we selected four such TFs from clusters 1 and 15, Ets96B, Helix loop helix protein 4C (HLH4C), Atonal (Ato), and Glass (Gl), and asked whether replacing the EcR motif with the motifs of these TFs would activate the enhancer in a TF-dependent but hormone-independent manner. This was indeed the case for all four TFs, and the effect was specific to the combination of motif and enhancer context (Fig. 2c), suggesting that these TFs might contribute regulatory functions equivalent to the activated EcR.

To assess the TF clusters independently of our approach, we asked whether they were enriched in Gene Ontology (GO) categories or in protein sequence features such as Pfam domains or short peptide motifs. Indeed, many such features were differentially distributed between the clusters (P < 0.01; empirical FDR = 0.1) and each cluster was enriched for at least one such feature (Fig. 2d and Extended Data Fig. 4; Supplementary Table 2). As expected, amino acid repeats known to mediate activation (for example, poly-glutamine14) or repression (for example, poly-alanine15) were enriched in activating clusters (1, 8) and repressing clusters (3, 5, 7), respectively. However, of several activating clusters, only cluster 1 was enriched in GO categories relating to development, suggesting a preferential use of cluster-1-type TFs during developmental gene regulation, which presumably relates to these TFs’ dependence on partner TFs, enabling combinatorial control. Similarly, only repressing cluster 7 but not 3 or 13 was enriched in GO categories relating to Notch signalling, cardiocyte differentiation, or morphogenesis, suggesting that repression might occur through various means that are differently employed in vivo. Indeed, the three repressing clusters also differed in the enrichment of peptide motifs known to bind the co-repressors C-terminal binding protein (CtBP; cluster 7) or Sin3A (cluster 3), suggesting a functional association between the TFs in these clusters and the respective co-repressors (see below).

These results show that the different TF clusters, obtained solely based on the TFs’ context-dependent regulatory functions, differ in several other aspects, which lends independent support to the clustering. It also suggests that the respective TFs are differentially employed in vivo (for example, during development), and that their distinct functions might arise through the recruitment of different types of cofactors (for example, CtBP versus Sin3A).

To assess the regulatory activities and the clustering in different cell types, we tested 171 TFs (9 to 17 TFs from each cluster) across six contexts in Kc167, BG3 and ovarian somatic cells derived from embryos, larvae and adult ovaries, respectively. These cell types differ increasingly from S2 cells in gene expression, enhancer activities, and the enhancers’ motif signatures3, yet TF activities were remarkably similar: all 18 pairwise comparisons had Pearson correlation coefficients (PCCs) ≥0.5 and 15 had PCCs ≥0.8 (all P < 1 × 10−3; Extended Data Fig. 5). Moreover, the TFs’ activity profiles across the 6 contexts support the original clustering in each cell type: pairs of TFs from within original clusters were more similar than pairs across clusters (≥1.4-fold and P < 1 × 10−6 for all four cell lines; Fig. 2e). Intrigued by these results, we tested the ability of 107 Drosophila TFs (90) and cofactors (17) to activate transcription in human HeLa cells and found a good quantitative agreement across all factors (PCC = 0.74; Extended Data Fig. 6). These results suggest that many TFs function predominantly, but not completely, independently of cell type. We note that alternative splicing and post-translational modifications (for example, downstream of cellular signalling pathways) probably alter and diversify the regulatory functions of individual TFs.

The regulatory functions of TFs are generally mediated through transcriptional cofactors, which typically lack DNA-binding domains and are recruited to enhancers by TFs. To assess whether cofactor functions are similarly diverse or potentially more uniform, we cloned 338 putative cofactors from diverse protein families (Fig. 3a and Supplementary Table 3) and tested their activating and repressing functions in all 24 contexts using GAL4-DBD-mediated recruitment as for TFs (Fig. 3b).

Figure 3: Transcriptional cofactors can be sufficient for activation and repression and have context-dependent regulatory functions.
figure 3

a, Pfam-domain content of the 338 putative cofactors tested. b, Scheme depicting enhancer complementation assays for cofactors. c, Overview of the regulatory activities for cofactors across all 24 contexts (bi-clustered 338 × 24 heat map) and zoom for selected cofactors. d, e, Diverse regulatory activities for bromodomain-containing cofactors (d) and subunits of the Mediator complex (e) (see Extended Data Fig. 7 for additional complexes).

PowerPoint slide

Full size image

Most cofactors (80%) were sufficient to activate or repress transcription in at least one context (≥1.5-fold; P < 0.05 after FDR correction for 24 × 338 tests), and the activities of well-studied factors matched their known functions (Fig. 3c–e): for example, P300 (also known as Nejire) strongly activated transcription in all contexts, as did the histone-methyltransferase Lost PHDs of Trr (Lpt) of the Set1/COMPASS-like complex, and the Mediator subunits MED15 and MED25, while the co-repressors CtBP, Sin3A and CoRest were strongly repressing in all contexts. Other cofactors had context-specific functions, including Chromator (Chro), TBP-associated factor 4 (Taf4) and Trithorax-related (Trr), which preferentially activated the housekeeping core promoter (Trr was even repressing in all other contexts). Chro is part of the non-specific lethal (NSL) complex which activates genes involved in cell proliferation and DNA replication16 and Taf4 is important at TATA-less promoters17. Similar to the corresponding TFs above, these cofactors might be part of a dedicated housekeeping regulatory program13.

Notably, different members of a single complex or domain family frequently had different regulatory functions (Fig. 3d, e and Extended Data Fig. 7), cautioning against the transfer of annotations based on these grounds. For example, the activities of bromodomain-containing (BRD) cofactors (Fig. 3d) range from the strongly activating P300 and CG7154, the orthologue of human BRD7 and BRD9, to the context-dependent CG30417, and to Polybromo that was strongly repressing in most contexts, consistent with its previous implication in transcriptional repression18. Similarly, Mediator subunits MED15 and MED25 were strongly activating in most contexts, MED23 and MED24 were context-dependent, and MED29 was repressing, consistent with the function of human MED2919 (Fig. 3e). This suggests that different TFs might interact with the Mediator complex through distinct subunits, or that complexes with variable composition and function might exist. Consistently, MED15 and MED25 interact directly with strong activators (for example, GAL4 and VP1620,21) and MED23 and MED24 are involved in signal-dependent and hormone-induced transcription, respectively22,23.

The activating and repressing effects across 24 contexts were highly similar for many TFs and cofactors (Fig. 4a), which provided a means to infer functional associations (Fig. 4b). As expected, P300, Lpt, MED15 and MED25 were assigned to globally activating TFs (cluster 8) and context-dependent cofactors to context-dependent TFs (for example, Mip120 and Bsh to cluster 10 and Chro to cluster 14). Interestingly however, the globally repressing cofactors Sin3A and CtBP were assigned to different clusters of repressing TFs (cluster 3 versus 7), in agreement with the differential enrichment of peptide motifs involved in Sin3A and CtBP recruitment (Fig. 2d). Many of the assignments are consistent with known physical interactions, including the interaction between Chro and Pzg24,25 or CtBP and Sna26. Indeed, the assignments were enriched for interactions reported in large-scale studies that used yeast two-hybrid assays27 or co-affinity purification25,28 (between 1.4- and 3.0-fold; all P ≤ 0.05; Fig. 4c). In addition, the human orthologues of TF–cofactor pairs interacted 1.8-fold (P = 0.025) more frequently than expected29. These results suggest that the TF–cofactor assignments reflect functional associations and predict that the enhancer activation obtained by TF recruitment should be boosted by the assigned but not by unrelated cofactors, even if the cofactors are not tethered via GAL4-DBD30 (Fig. 4d). Indeed, in this experimental setup30, the activation by Clk-recruitment (cluster 8) was further boosted by increasing amounts of untagged P300, but not by Chro or Tbp. In contrast, Chro specifically boosted CG17186 (cluster 14) and Tbp specifically boosted Bsh (cluster 10), consistent with the respective assignments.

Figure 4: TF–cofactor assignments through functional similarities.
figure 4

a, TFs (ovals) and cofactors (diamonds) show similar regulatory activities and localize next to one another in a bi-clustered 812 × 24 heat map (selection shown; see Supplementary Table 1 for all activities). Coloured as in Fig. 3c–e. b, TF clusters (as in Fig. 2a) and assigned cofactors. Highlighted are clusters 8, 14 and 3, 5, 7 and assigned cofactors. c, TF–cofactor assignments are enriched for physical interactions (red bars) compared to shuffled assignments, for which the box-plots indicate the 10th, 25th, 50th, 75th and 90th percentiles; *hypergeometric P ≤ 0.05. d, Boosting of Clk (top), CG17186 (middle) and Bsh (bottom) induced enhancer activities by untagged P300 (left), Chro (centre) and Tbp (right). Error bars show standard deviations (n = 4, biological replicates). **P < 1 × 10−2; ***P < 1 × 10−3.

PowerPoint slide

Full size image

Enhancer complementation assays provide a unique annotation and categorization of TFs and cofactors based on their regulatory functions, independent of the factors’ endogenous roles and complementary to previous classifications through genetics, sequence comparisons, or genomics (for example, ChIP-seq). For many factors, including 266 putative TFs and cofactors (‘CG’ genes; Extended Data Fig. 8), our work provides the first functional characterization. All data are available at http://factors.starklab.org.

The existence of equivalence groups among TFs and cofactors with diverse context-dependent functions, even amid activators and repressors, has profound implications for our understanding of transcriptional gene regulation: while some enhancers might be controlled predominantly by individual activators, others may rely on specific combinations of distinct regulatory functions that are complementary and each insufficient for activation. It is therefore conceivable that different types of enhancers are controlled through non-overlapping sets of TFs and cofactors, enabling separate transcriptional programs even within individual cells (for example, ref. 13). The approach and categorization presented here provide a framework to dissect the molecular and biochemical nature of these functions and the mechanisms by which cooperativity at enhancers is established and transcriptional activation of target core promoters is achieved. Understanding these mechanisms will be crucial at a time when enhancer function and its control by TFs and cofactors are becoming increasingly central to our understanding of gene regulation in development and disease and the focus of novel therapeutic strategies.

Methods

No statistical methods were used to predetermine sample size. The experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment.

Cloning of N-terminal GAL4-DBD-tagged TF and cofactor library

For TFs, gateway-compatible entry clones (Invitrogen) containing the open reading frames (ORFs) lacking stop codons were obtained from ref. 11. Drosophila Act5C-promoter driven expression clones were created using the Gateway system. The TF ORFs were shuttled into the GAL4-DNA binding domain (DBD) containing destination vector pAGW-GAL4-DBD (cloned as described below) by mixing 100 ng of TF entry clone, 100 ng of pAGW-GAL4-DBD and 0.7 μl of LR clonase II enzyme mix (Invitrogen). The identities of all TF entry clones have been confirmed by Sanger sequencing using the primers 5′-CCCAGTCACGACGTTG-3′ and 5′-CACAGGAAACAGCTATG-3′. Note that we tested the full-length transcription factors, including their DBDs, as trans-activating and DNA-binding functions might not always reside in entirely separate protein domains. While this implies that the fusion proteins might bind via the TFs’ DBDs in addition to the GAL4-DBD mediated recruitment, this does not influence the results of the assay: the assay itself measures transcriptional activation independently of where TF binding occurs and we expect that the TFs’ DBDs have at most minor effects on binding strengths as the GAL4-DBD binds to DNA already very strongly.

For cofactors, we compiled a list of 338 cofactors based on several criteria. We included proteins containing Pfam domains typical for transcriptional cofactors (for example, HAT, HDAC, SET, Chromo, Bromo), proteins which are part of chromatin modifying or remodelling complexes or part of complexes associated with RNA polymerases (for example, SAGA, Polycomb, TFIID, Mediator), and Drosophila proteins which are homologues of mammalian chromatin-associated proteins (Supplementary Table 3). We amplified the cofactor ORFs from cDNA using oligonucleotides containing Gateway-compatible attB-sites (5′-GGGGACAAGTTTGTACAAAAAAGCAGGCTTC-3′ and 5′-GGGACCACTTTGTACAAGAAAGCTGGGTC-3′) for subsequent entry clone creation. The primer sequences have been chosen to be as close as possible to an annealing temperature of 60 °C which we calculated using the formula with cA, cT, cG and cC being the number of adenines, thymines, guanines and cytosines, respectively. The full list of resulting primer sequences (lacking the attB sequences) is listed in Supplementary Table 3; for 18 of the cofactors no primer sequences are available because we obtained these entry clones from ref. 11 categorized as TFs but manually re-categorized them as cofactors based on their annotation in FlyBase31 or their protein domain content32. For cDNA generation, RNA was isolated from S2 cells and reverse transcribed as described in ref. 33. For PCR amplification, KOD and KOD XL DNA (Merck Millipore) and KAPA HiFi (KAPA) polymerases were used according to manufacturer’s specifications. We created Gateway entry clones by mixing 1 μl of PCR reaction, 100 ng of pDONR221, and 1 μl of BP clonase II enzyme mix (Invitrogen). The identities and correctness of all entry clones have been ensured using Sanger and next-generation sequencing (see below) and we deposited them at Addgene (http://www.addgene.org/Alexander_Stark/). The cofactor ORFs were then shuttled to the Drosophila Act5C-promoter driven destination vector pAGW-GAL4-DBD as described for TFs.

Verification of cofactor clones by Sanger and next-generation sequencing

The insert flanks of all obtained cofactor entry clones have been Sanger-sequenced and automatically checked to cover the TSS and TTS of one of the isoforms annotated by FlyBase. All entry clones passing additional manual visual inspection using BLAT and the UCSC genome browser have been subjected to further verification by next-generation sequencing as follows. A pool of 100–300 entry clones corresponding to a total of 5 μg DNA solved in 50 μl TE buffer was sonicated (duty cycle, 20%; intensity, 5; 200 cycles per burst; time, 90 s) to 200–400 bp using a S220 Focused-ultrasonicator (Covaris) as described in ref. 11. The fragmented plasmid pool was then prepared for deep sequencing using the Illumina DNA Sample Prep kit and sequenced using a HiSeq2000 (Illumina) producing 50-nt reads. The resulting reads have been assembled and analysed using PrInSes-C34. All insert sequences not starting with ATG, containing a stop codon or a frameshift were immediately rejected. All sequences with less than five mutations leading to non-synonymous amino acid changes were immediately accepted. The remaining sequences were translated, aligned against the respective protein sequence, and manually decided. The next-generation sequencing reads have been deposited at the NCBI Sequence Read Archive (SRA) under the accession SRS806429; the PrInSes-C-generated full-length transcript sequences are available at http://factors.starklab.org and in Supplementary Data 1, and the cofactor Gateway entry clones from Addgene (http://www.addgene.org/Alexander_Stark/).

Cloning of destination vector pAGW-GAL4-DBD

We cloned a destination vector to conveniently create vectors expressing N-terminally V5- and GAL4-DBD-tagged TFs and cofactors under the control of the Drosophila Act5C promoter using the Gateway cloning system. pAGW-GAL4-DBD was cloned by amplifying the GAL4-DBD from pBPGUw35 using one oligonucleotide containing the V5-tag (peptide sequence MGKPIPNPLLGLDST) 5′-TCTGATATCATGGGGAAGCCAATCCCTAATCCCCTTCTGGGACTCGACTCTACCGGCGGCTCTATGAAGCTACTGTCTTCTATCGAACA-3′ and the oligonucleotide 5′-TATACCGGTGGCCGCCGCCCGACGATACAGTCAACTGTCTTTGAC-3′. Amplification was performed using KOD Polymerase (Merck Millipore) according to the manufacturer’s instructions. The resulting PCR product was digested using EcoRV and AgeI and ligated into pAGW (Drosophila Gateway Vector Collection), which was digested using the same enzymes, thereby replacing eGFP with V5-GAL4-DBD.

Cloning of luciferase reporter vectors

We created Gateway-compatible (Invitrogen) destination vectors to conveniently clone reporter vectors for different regulatory contexts based on firefly luciferase transcribed from a housekeeping core promoter (hkCP; promoter of ribosomal gene RpS1213) or a developmental core promoter (dCP; Drosophila synthetic core promoter (DSCP) derived from Eve35).

We created the destination vector attR_dCP_luc by digesting pGL4.26 (Promega) with FseI and BglII and ligating a fragment containing DSCP and luc+, thereby replacing the minimal promoter and luc2 with DSCP-luc+. We digested the resulting vector with KpnI and BglII and ligated a fragment containing the attR Gateway cassette, yielding attR_dCP_luc. We created two hkCP-driven destination vectors containing a Gateway cassette either upstream (attR_hkCP_luc) or downstream (hkCP_luc_attR) of the luciferase reporter gene by using the plasmid pGL3 (Promega) as a basis and replacing the SV40 promoter with the promoter of RpS12 as described in ref. 13. The resulting vector was digested using either KpnI and BglII (to create attR_hkCP_luc) or AfeI (to create hkCP_luc_attR); in both cases, we amplified a Gateway attR cassette using oligonucleotides containing the respective restriction sites, and digested and ligated it into the digested plasmid.

All enhancers, motif mutant contexts and other motif or backbone mutant variants were either PCR amplified with primers containing attB Gateway sites or ordered as synthesized fragments (IDT), shuttled into entry clones using TOPO or BP Clonase II (both Invitrogen), and shuttled into the luciferase destination vectors using the LR clonase II enzyme mix (Invitrogen) by mixing 1 μl of PCR product or synthesized DNA solved in TE buffer, 100 ng of destination vector and 0.7 μl of LR clonase II enzyme mix (Invitrogen).

We used a modified version of pRL-TK (Promega) to normalize the firefly signal for transfection efficiency and cell number. Ubi-RL has been created by cloning a region upstream of the gene Ubi-p63E (chr3L: 3901760-3902637) upstream of the Renilla luciferase gene in reverse orientation using NheI and BglII.

Drosophila cell culture

S2 cells, derived from embryos36, were obtained from Life Technologies and grown in Schneider’s Drosophila Medium (Life Technologies 21720-024) supplemented with 10% FBS (Sigma F7524) and 1% penicillin/streptomycin (Life Technologies 15140-122) grown in T75 flasks (ThermoScientific 156499) at 27 °C and passaged every 2–4 days. BG3 neuroblast-like cells, derived from larvae37, were obtained from the Drosophila Genomics Resource Center (DGRC) and grown in Schneider’s Drosophila Medium supplemented with 10% FBS, 1% penicillin/streptomycin, and 10 μg ml−1 Insulin (Sigma-Aldrich I1882) in T75 flasks at 27 °C and passaged every 3–4 days. Kc167 cells, derived from embryos38, were obtained from DGRC and grown in M3/BPYE Medium containing 5% FBS and 1% penicillin/streptomycin in T75 flasks at 27 °C and passaged every 2–3 days. Ovarian somatic cells (OSCs), derived from adult ovaries39, were obtained from the laboratory of J. Brennecke and grown in Shields and Sang M3 Insect Medium (Sigma-Aldrich S8398) supplemented with 10% FBS, 1% insulin, 1% glutathione, 1% fly extract, and 1% penicillin/streptomycin in T75 flasks at 27 °C and passaged every 2–3 days. All cell lines used are regularly checked for mycoplasma contamination.

Transfections of Drosophila cell lines

S2 cell transfections were performed using jetPEI (peqlab 13-101-40N). Four hours before transfection, 30,000 cells (30 μl of a 106 cells per ml suspension) were seeded in clear polystyrene 384-well plates (ThermoScientific 164688). For each transfection, we used 30 ng firefly luciferase reporter plasmid, 3 ng Renilla luciferase expressing plasmid Ubi-RL, and 3 ng GAL4-DBD-TF/cofactor or GAL4-DBD-GFP fusion protein expressing plasmid. Beforehand, we assayed the effects of using different amounts of GAL4-DBD fusion protein expressing plasmid and chose 3 ng (Extended Data Fig. 9). The DNA solution containing 36 ng DNA in 5 μl TE buffer was filled up to 15 μl using sterile 150 mM NaCl (polyplus) and prepared in 96-well plates. Transfection reagent (15 μl total: 13.95 μl 150 mM NaCl, 1.05 μl jetPEI) was added to each well of the 96-well plates and mixed rigorously. After 30 min incubation at 25 °C, cells were transfected in quadruplicates by transferring each transfection mix four times (6 μl each) to four adjacent wells of a 384-well plate containing the seeded cells. Luciferase assays were performed after 48 h of growth at 27 °C. Handling the transfection mixes and all subsequent pipetting steps have been performed using a Bravo Automated Liquid Handling Platform (Agilent). Kc167, BG3, and OSC cell transfections were performed using jetPEI in the same way as described above for S2 cells with the exception of transfection reagent composition: 15 μl total containing 14.1 μl 150 mM NaCl and 0.9 μl jetPEI.

HeLa cell culture and transfections

Human HeLa cells (gift from the laboratory of J. M. Peters) were grown in DMEM medium (Gibco 52100-047) supplemented with 10% heat-inactivated FBS, 1% penicillin/streptomycin and 2 mM L-glutamine (Sigma G7513) in T75 flasks at 37 °C in an atmosphere of 95% air and 5% carbon dioxide. All cell lines used are regularly checked for mycoplasma contamination. We performed HeLa cell transfections using a self-prepared 1 mg ml−1 PEI (25,000 MW, Polysciences 23966) stock solution in PBS (pH adjusted to pH 4.5 and sterile filtered). On the day before transfection we seeded 30 μl of a suspension containing 4,000 HeLa cells in medium (DMEM, 10% FBS, penicillin/streptomycin) into each well of a 384-well plate. Three microlitres of a PEI/DMEM mix (0.24 μl PEI filled to a total of 4.5 μl using DMEM without FBS and penicillin/streptomycin and incubated at room temperature for 5 min) were added to 3 μl of a DNA/DMEM mix (44.5 ng firefly luciferase reporter vector, 4.45 ng TF expression vector (created using pAGW-CMV_ GAL4-DBD, see below) and 4.45 ng pRL-CMV vector for transfection normalization (Promega #E2261) in DMEM without FBS and penicillin/streptomycin. The resulting DNA/PEI mix in DMEM was incubated at room temperature for 30 min and subsequently added to the seeded cells. We performed cell lysis and luciferase assays using the Promega dual-luciferase reporter assay system (Promega E1910) according to the manual.

We created the Gateway destination vector pAGW-CMV_GAL4-DBD by replacing the Drosophila Act5C promoter in pAGW-GAL4-DBD with a region containing the CMV enhancer and the T7 promoter amplified from pRL-CMV using the primers 5′-CGACAGATCTTCAATATTGGCCATTAGCCATAT-3′ and 5′-GGTGGCTAGCCTATAGTGAGTCGTATTA-3′.

Luciferase assays

Dual-luciferase assays were performed using self-prepared substrate solutions (D-Luciferin and Coelenterazine have been obtained from GoldBio LUCK-250 and pjk-Gmbh 102111) and lysis buffer as described in ref. 40. For cell lysis, the supernatant was removed and 30 μl of lysis buffer added and incubated gently shaking for 30 min. Ten microlitres of the cell lysates were transferred to black 384-well plates for luminescence assays (Nunc MaxiSorp, Sigma-Aldrich P6491-1CS). All pipetting steps have been performed using a Bravo Automated Liquid Handling Platform (Agilent). Luminescence was measured after adding 20 μl of each substrate, for firefly and Renilla luciferase respectively, using a Biotek Synergy H1 plate reader coupled to a plate stacker.

Luciferase data analysis and plots

We normalized all firefly luciferase signals to the signal of Renilla luciferase to control for transfection efficiency and cell number (the relative luciferase signal). We then further normalized all relative luciferase signals for TF- and cofactor-GAL4-DBD transfections to relative luciferase signals obtained for GAL4-DBD-GFP transfections (fold-change over GFP). We assessed statistical significance by two-sided unpaired t-tests on the two sets of quadruplicate relative luciferase signals (GAL4-DBD-TF/COF versus GAL4-DBD-GFP). Throughout the paper, ‘activation’ was defined as a fold-change ≥1.5 (P < 0.05), and ‘repression’ was defined as a fold-change ≤1/1.5 (P < 0.05), both compared to the signal for GAL4-DBD-GFP. We corrected the P values for multiple testing using the Benjamini and Hochberg method as implemented in R (p.adjust with method ‘BH’ or its alias ‘fdr’). All statistical calculations and graphical displays, if not stated otherwise, have been performed using version 2.15.3 of the R software suite41.

TF cluster feature enrichment analysis

Enrichment analyses have been performed for each of the 15 clusters and for 6 types of features. To first obtain a coarse functional characterization of the clusters, we assessed the enrichments and depletions of TFs which are able to activate or repress a developmental (dCP) or housekeeping (hkCP) core promoter on their own (≥1.5-fold activation or repression (P < 0.05), both compared to the signal for GAL4-DBD-GFP when tested on a context comprised of UAS sites upstream of a developmental core promoter (4×UAS-dCP) or a housekeeping core promoter hkCP (4×UAS upstream hkCP)). Homopolymeric amino acid repeat motifs have been de novo discovered using MEME42 (version 4.8.1, q-value threshold of 1 × 10−5) in TFs that activated or repressed on their own outside enhancer contexts (tested in the 4×UAS dCP context; ≥1.5-fold; P < 0.05). Pfam domain32 signature matches in the Drosophila proteome have been generated using hmmer43 (version 3.0b3, e-value threshold of 0.01). Eukaryotic Linear Motifs44 (ELM; version 08/2014) were matched to the amino acid sequences of the tested TF protein isoforms, after masking the TFs’ Pfam. Additionally, Gene Ontology45 (GO) annotations, and gene expression patterns in the Drosophila embryo as annotated by ref. 46 (IMAGO) have been subjected to enrichment and depletion analyses.

To control for multiple testing, we empirically determined false-discovery rates (FDRs) for the different hypergeometric P values. For this, we repeated the feature enrichment analyses 1,000 times, each after randomly shuffling the TF-to-cluster assignments, and recorded the best (that is, most significant) P values. We then adjusted the original P values such that only 10% of the 1,000 random controls reached the P values of the original data (FDR < 10%). Following this protocol, we separately adjusted the FDR cut-off for each cluster (15) and feature type (ELM, MEME, Pfam, GO, IMAGO).

Validation with the TFs’ endogenous motifs

To assess if tethering via the GAL4-DBD reflects the different TFs’ regulatory functions when bound to their endogenous motifs, we selected two sets of TFs, three TFs that preferentially activated the CGCG- versus the GATA-context (Fig. 1e) and four TFs that preferentially activated the hormone-receptor contexts; Fig. 2c). We replaced each UAS site in the enhancer mutant contexts S2-1 CGCG, S2-1 GATA, and Nhe2 EcR3,12 (which also corresponds to an endogenous TF motif in the wild-type enhancers, for example, the EcR motif for the hormone contexts) with a sequence corresponding to the consensus motif of the respective TF as reported in refs 47, 48. (Dfd: CTTAATGA, Hey: CAGCCGACACGTGCCCC, Ets21C: ATTTCCGGT, Ato: AACAGGTGG, Ets96B: ACCGGAAGTAC, Gl: ATTTCAAGAATA, HLH4C: AAAAACACCTGCGCC). The enhancer rescue constructs were synthesized by IDT, shuttled into the luciferase reporter vector attR_dCP_luc using the Gateway system and tested in luciferase assays in S2 cells exactly as described above.

TF–cofactor association assays

To assess potential functional associations of assigned TFs and cofactors, we followed the strategy from ref. 30, recruiting TFs via GAL4-DBD and providing untagged cofactors. For this, we chose contexts in which the different TFs (Clk of cluster 8, Bsh of cluster 10, and CG17186 of cluster 14) were active (4×UAS-dCP for Clk and 4×UAS-upstream-hkCP for Bsh and CG17186). We prepared DNA mixes to be transfected containing 29 ng firefly luciferase reporter plasmid, 3 ng Renilla luciferase expressing plasmid Ubi-RL, 1 ng (Bsh and CG17186) or 0.5 ng (Clk) of GAL4-DBD–TF fusion protein expressing plasmid and an increasing series of untagged cofactor expressing plasmid (0 ng, 0.003 ng, 0.006 ng, 0.012 ng, 0.023 ng, 0.047 ng, 0.094 ng, 0.188 ng, 0.375 ng, 0.75 ng, 1.5 ng, 3 ng). We kept the total amount of transfected plasmid DNA constant at 36 ng for all experiments using a GFP-expressing plasmid. To clone the expression plasmids for the untagged cofactors and GFP, we used the Gateway-compatible vector pAW (Drosophila Gateway Vector Collection). The remaining experimental procedure and analysis was performed as described above.

Transcription factor clustering, visualization, and assignment of cofactors to transcription factors

We clustered the 474 TFs based on the log2-transformed fold-change values (TF over GFP) from all 24 contexts. First, we standardized all contexts and constructed a k-nearest-neighbour graph (k = 15). We used the Euclidean distance as distance measure as it reflects both the variation of the enhancer activity profile across contexts and the effect sizes within each context; that is, it is able to discriminate between strong and weak activators and repressors even if they vary similarly across the 24 contexts. Next, we took a symmetrized (A + AT) adjacency matrix of this graph and solved multiclass spectral clustering as described in ref. 49 and implemented in the Python package scikit-learn50. In order to decide about the number of clusters and to assess the clustering validity, we analysed the clustering stability upon bootstrapping the data set51. In order to visualize the data, we mapped the data onto a plane by a specialized nonlinear dimensionality reduction technique (t-SNE)52. The algorithm provides the visualization by mapping data points close in the original space to nearby locations in the plane, preserving the local structure. We extended the k-nearest-neighbour graph to include cofactors by comparing the log2-transformed fold-change values (cofactor over GFP) of cofactors and TFs (k = 5, Euclidean distance). The locations of the cofactors in the visualization were obtained from spring layout.

TF candidate recovery of enhancer mutants

We know that UAS sites in the enhancer mutant contexts most probably replace binding sites that are functional3 but we do not know which TFs bind them in vivo. In order to check whether we recover these positive controls in the enhancer mutants, we took all the TFs expressed in S2 cells (RPKM > 1) (ref. 53) for which motifs are known3. We scanned the wild-type enhancer sequences (S2-1-wt, S2-2-wt, S2-3-wt, Ubi-1-wt, Ubi-2-wt, Ubi-3-wt) for motif matches with P < 9.76 × 10−4 (1/4,096) using an in-house motif-detection program. For each mutant context, we considered only those TFs for which any of its motif matches had at least 5 mutated base pairs. In the resulting set of TFs (Extended Data Table 1) there is at least one TF per each of the enhancer mutant contexts that activated the respective context when recruited via the GAL4-DBD (≥1.5-fold activation compared to GFP; P < 0.05).

Cell type analysis—distances intra- versus inter-cluster

We tested a subset of the original 472 TFs in four different cell types (S2, Kc167, BG3 and OSC). This subset consists of 171 TFs covering all the 15 clusters by 9–17 TFs, including all the TFs mentioned in the main text. In each cell type, we computed Euclidean distances after standardizing the log2-transformed fold-change values in each context. Then we compared the distances of intra-cluster TF–TF pairs (both TFs belong to the same cluster) to inter-cluster TF–TF pairs (each of the TFs belongs to a different cluster). In order to test whether the medians of these two groups of distances are significantly different, we determined empirical P values as follows. We randomly shuffled TF-to-cluster assignments 106 times and each time computed the medians of the distances for both groups. We mark the P values P < 1 × 10−6 as we never obtained a difference between the medians of intra- and inter-cluster distances as large as for the actual data for any of the cell types.