A compendium of Caenorhabditis elegans regulatory transcription factors: a resource for mapping transcription regulatory networks

Transcription regulatory networks are composed of interactions between transcription factors and their target genes. Whereas unicellular networks have been studied extensively, metazoan transcription regulatory networks remain largely unexplored. Caenorhabditis elegans provides a powerful model to study such metazoan networks because its genome is completely sequenced and many functional genomic tools are available. While C. elegans gene predictions have undergone continuous refinement, this is not true for the annotation of functional transcription factors. The comprehensive identification of transcription factors is essential for the systematic mapping of transcription regulatory networks because it enables the creation of physical transcription factor resources that can be used in assays to map interactions between transcription factors and their target genes. By computational searches and extensive manual curation, we have identified a compendium of 934 transcription factor genes (referred to as wTF2.0). We find that manual curation drastically reduces the number of both false positive and false negative transcription factor predictions. We discuss how transcription factor splice variants and dimer formation may affect the total number of functional transcription factors. In contrast to mouse transcription factor genes, we find that C. elegans transcription factor genes do not undergo significantly more splicing than other genes. This difference may contribute to differences in organism complexity. We identify candidate redundant worm transcription factor genes and orthologous worm and human transcription factor pairs. Finally, we discuss how wTF2.0 can be used together with physical transcription factor clone resources to facilitate the systematic mapping of C. elegans transcription regulatory networks. wTF2.0 provides a starting point to decipher the transcription regulatory networks that control metazoan development and function.


Background
Metazoan genomes contain thousands of predicted proteincoding genes. During development, pathology, and in response to environmental changes, each of these genes is expressed in different cells, at different times and at different levels. Spatial and temporal gene expression is controlled transcriptionally through the action of regulatory transcription factors (TFs) [1,2]. Transcription of each gene can be upor down-regulated by TFs that bind to cis-regulatory DNA elements. These elements include upstream elements located in the proximal promoter, and enhancers or silencers that can be located at a greater distance from the transcription start site. Frequently, the expression level of a gene is the result of a balance between transcription activation and repression governed by multiple cis-regulatory elements and, hence, multiple TFs. The combinatorial nature of gene transcription provides an exquisite level of flexibility to regulate genome expression.
The understanding of differential gene expression at a genome-wide, or systems, level has been greatly facilitated by the mapping and analysis of transcription regulatory networks ( Figure 1) [3][4][5]. Such networks are composed of two types of components, or nodes: the gene targets that are subject to transcriptional control and the TF proteins that execute transcriptional control. Whereas transcription regulatory networks have been extensively studied in relatively simple unicellular systems, they remain largely unexplored in complex, metazoan systems.
The nematode Caenorhabditis elegans is a powerful model to decipher metazoan transcription regulatory networks. The C. elegans genome has been completely sequenced and is predicted to contain 19,735 protein-coding genes (WormBase WS140) [6]. Several functional genomic resources enable the systematic dissection of differential gene expression at a systems level and in a high-throughput manner. For instance, microarrays are available to investigate temporal and, to a certain extent, spatial gene expression levels [7][8][9]. In addition, C. elegans 'ORFeome' [10] and 'Promoterome' [11] resources provide open reading frame (ORF) and promoter clones, respectively. These clones can be used for a wide variety of experiments that aim to dissect transcription regulatory networks (see below).
With seven years of progressive refinement of the genome annotation since the publication of the genome sequence [12][13][14], comprehensive predictions of protein-coding genes are available. However, there is no up-to-date compendium of predicted C. elegans TFs. Several lists of putative C. elegans TFs have been generated previously (Table 1; we refer to this combined set as wTF1.0 (worm transcription factors version 1.0)), but none of these are comprehensive or readily accessible. The earliest lists were created by scanning sequences of C. elegans proteins, as predicted from the genome annotation, for well-defined DNA binding domains: Hobert and Ruvkun [15] focused on homeodomain, paired domain, T-box, basic helix-loop-helix (bHLH), basic region leucine zipper (bZIP), Fork Head, erythroblast transformation specific (ETS) and nuclear hormone receptor (NHR) proteins; and Clarke and Berg [16] focused on various zinc-finger proteins. For comparative genomic studies primarily focusing on Drosophila melanogaster [17] or Arabidopsis thaliana [18], a list of predicted C. elegans proteins containing various DNA binding domains was compiled. While several (sub)-families of C. elegans TFs have been studied in greater detail (for example, bHLH [19], CUT homeodomain [20], and DM zinc-finger [21]), the most recent list of "all C. elegans TFs" was compiled five years ago when Riechmann and colleagues [18] scanned WormPep 20 (19,101 proteins). During the past few years, the creation of improved computational tools [13] and the completion of the C. briggsae genome sequence [14] have enabled a great improvement in the annotation of the C. elegans genome. Here, we used a combination of bioinformatics and extensive manual curation to generate wTF2.0, a comprehensive compendium of predicted C. elegans TF genes. We discuss how wTF2.0 can be used together with physical ORFeome and Promoterome clone resources to decipher transcription regulatory networks that control metazoan differential gene expression at a systems level.

Results and discussion TF predictions: Gene Ontology term-based searches
To identify a comprehensive compendium of predicted worm TFs, we first interrogated WormBase version 140 (WS140) [6] for proteins that possess domains annotated with one of the following Gene Ontology (GO) terms: 'regulation of transcription, DNA-dependent', 'transcription factor activity', and 'DNA binding'. (WS140 is the most recent reference release of WormBase that is permanently accessible.) We identified a total of 930 proteins ( Figure 2, Additional data file 1). Of these, 232 were identified by all three GO terms, 368 by two GO terms and 330 by only one GO term ( Figure 3). We observed that this collection of proteins not only contains predicted regulatory TFs, but also proteins that function in other nuclear processes (for example, DNA replication and repair). Moreover, it contains numerous false positive predictions (for example, small GTP-binding proteins). We removed both types of false positives (Figure 2, Additional data file 2). Next, we examined each of the remaining proteins for the presence of a predicted DNA binding domain either by visual inspection of the protein sequence (AT-hooks and C2H2 zinc-fingers), or using InterPro v10.0 (2005) [22], SMART [23] and Pfam [24] databases (Additional data file 2). We found several proteins that, upon closer inspection, do not possess a DNA binding domain despite their WormBase protein domain annotation (Additional data file 2). For instance, several proteins were found that are annotated to be a NHR TF, but that only contain a predicted ligand binding domain. However, we retained proteins that do not have a clear DNA binding domain but for which experimental evidence is avail- able that supports their function as a TF. For example, we included SKN-1, a bZIP protein known to bind DNA in a sequence-specific manner [25]. In total, 369 proteins (40%) were removed (Figure 2a, Additional data file 2). As expected, combining all three GO terms was the most robust method for identifying predicted TFs, as 96% of these proteins were retained. The GO term DNA binding by itself was least robust as only 16% of these were retained. However, this can readily be explained by the retrieval of proteins that do bind DNA but that are not involved in transcriptional regulation.
Additional data file 1 wTF2.0: a collection of predicted C. elegans transcription factors wTF2.0: a collection of predicted C. elegans transcription factors Click here for file Additional data file 2 Overview of manually curated genes that were left out of wTF2.0 Overview of manually curated genes that were left out of wTF2.0 Click here for file

TF predictions: DNA binding domains
Upon examination of the remaining 561 proteins, we noticed that several well known TF families were underrepresented compared to wTF1.0, or even absent (for example, bHLH, C2H2 zinc-fingers and MADF (Mothers Against Dpp Factor)). This suggests that the predictions based on GO annotations alone suffer from a high false negative rate. To address this issue, we searched WormBase for each protein domain known to be involved in sequence specific DNA binding (Table 1). In addition, we added several TFs found by yeast one-hybrid assays (for example, TFs containing RPEL and FLYWCH domains [26] (data not shown)). We used visual inspection (C2H2 zinc-fingers and AT-hooks), InterPro, SMART and Pfam to verify these predictions and, in total, added 369 additional, putative TFs to the compendium. Finally, we added 4 proteins: 3 of which are homologs of known mammalian TFs (BAR-1, HMP-2 and WRM-1, homologs of mammalian β-catenin) and one that has been described in the literature (SDC-2 [27]). In total, amongst the 19,735 predicted protein-coding genes, we identified 934 predicted C. elegans TF genes (Additional data file 1). Taken together, the combination of computational queries and manual curation results in a comprehensive compendium of C. elegans TF-encoding genes. We refer to this compendium as wTF2.0. . Future studies will determine if and how these domains function together in DNA binding specificity and, consequently, target gene selection. Interestingly, most human orthologs of these TFs also contain multiple, distinct DNA binding domains (Additional data file 3), indicating that the occurrence of multiple DNA binding domains in a TF is not worm specific. Comparison of wTF2.0 to wTF1.0 revealed that, of 48 TF families, 23 are unique to wTF2.0 (Table 1). This is likely because for the collective predictions in wTF1.0, only the major DNA binding domains were included. In addition, some protein domains have only recently been annotated to function in DNA-binding, including SAND [28], and THAP zinc-finger [29] domains. For several of the domains unique to wTF2.0 (CP2, WH-DAC, IPT/ TIG, TEA/ATTS, WT1, YL1), the genes encoding them were not actually annotated until after WormPep 20 and could, therefore, not have been included in the wTF1.0 collections.  improvements in gene-prediction software. We have noted several changes in gene annotations in WormBase releases subsequent to WS140 that affect wTF2.0 (Additional data file 4). There have been additions, such as Y55F3AM.7 (created in WS146 and encoding a C2H2 zinc-finger protein), and eliminations, such as Y60A9.2 (encoding a CCCH zinc-finger protein in WS140, but designated a pseudogene since WS141).

TF families
We have also noted more subtle adjustments in gene structure based on TWINSCAN [13] suggestions of different splicing patterns. For example, a modified gene structure for F22A3.5 that meant the gene product would then include a complete homeodomain was adopted in WS143, with support from C. briggsae genome sequence. Similar gene structure changes that would lead to intact homeodomains for F34D6.2, R04A9.5 and ceh-31, and an intact bHLH domain for hlh-19 (see comments in Additional data file 1) may yet be incorporated into WormBase. Taken together, we expect that wTF2.0 will be a dynamic resource but that a relatively small number of TFs will be removed and added over time.  [30] observed that mouse TF genes are more likely to undergo alternative splicing than other mouse genes (62% compared to 29%). These alternatively spliced TF genes may yield functionally different TFs that may bind DNA with different specificities and affinities and, as a consequence, regulate different sets of target genes. In contrast, the percentage of C. elegans TF genes that undergo alternative splicing is only slightly higher than the percentage of all protein-coding genes that are alternatively spliced (15% versus 10%) [31] (this study). This observation suggests that higher percentages of TF gene splicing may contribute to increased organism complexity. Finally, it is important to note that several C. elegans TFs can be expressed from multiple alternative promoters (Additional data file 5). Alternative promoters are likely to drive different patterns and levels of TF production, which may contribute to the complexity of combinatorial gene expression.
Additional data file 5 Alternative splice forms and promoters Alternative splice forms and promoters Click here for file

Functional TFs: dimers
Several TFs, including bHLH, NHR and bZIP proteins, are known to bind DNA as either homo-or heterodimers, and the different dimer combinations that occur in vivo determine the actual number of TF complexes. For instance, the mini-Transcription regulatory networks provide models to understand differential gene expression at a systems level Figure 1 Transcription regulatory networks provide models to understand differential gene expression at a systems level. Transcription regulatory networks are composed of two types of components, or nodes: the genes involved in the system and the TFs that regulate their expression. Protein-protein interactions between TFs and protein-DNA interactions between TFs and their target genes can be visualized in transcription regulatory networks. The dashed line represents TF-TF protein-protein interaction (heterodimer). Arrows represent protein DNA interactions that result in transcription activation; the blunt 'arrow' represents protein-DNA interaction that results in repression of transcription.
Gene X A X mum number of TF complexes would be half the total number of TFs predicted if each TF were to exclusively dimerize with one other TF. Alternatively, the total number of functional TFs could be much larger than the number of predicted TF genes if each TF dimerizes with multiple other TFs. To start addressing this issue, we retrieved a network of TF-TF interactions that were identified in a large-scale yeast two-hybridbased protein-protein interaction mapping study [32] (Figure  4, Additional data file 6). We found 68 putative TF-TF dimers involving 71 TFs: 35 between members of different TF families and 33 between members of the same TF family. Of these 33, 7 are putative homodimers and the remaining 26 are putative heterodimers. Interestingly, the TF dimerization network suggests that certain TFs, such as NHR-49, can function as dimerization hubs. NHR-49 is involved in the regulation of fat storage and life span [33], but it is not known if NHR-49 functions in these processes as a homodimer or in concert with other NHR TFs. It is noted that the current TF dimerization network is only a small representation of all TF dimers. This is because some TFs may only form dimers on their cognate DNA and may, therefore, not be detected by yeast two-hybrid assays; and because the current worm 'interactome' (WI5) only contains approximately 5% of all protein-protein interactions that can be detected by yeast two-hybrid assays [32]. Future systematic TF-TF proteinprotein interaction mapping projects are required to determine the total complement of TF dimers. Although it is difficult to interpret interactions between TFs from different families, they could point to putative combinatorial regulation of target genes. Taken together, assuming that many TFs can function both as monomers and dimers, the number of functional TFs will likely exceed the number of predicted individual TF proteins.
Additional data file 6 Overview of protein interactions involving wTF2.0 TFs Overview of protein interactions involving wTF2.0 TFs Click here for file

TF families: redundancy
For the systematic mapping of transcription regulatory networks, it is important to identify redundancy between closely related TF genes. This is because redundant genes have similar, overlapping or identical biological functions and, thus, results obtained with an individual TF may be difficult to   [34], the GATA factors med-1 and med-2, and end-1 and end-3 [35], and the T-box genes tbx-8 and tbx-9 [36], and tbx-37 and tbx-38 [37]. To identify additional putative redundant or highly similar TF genes, we used ClustalX analysis to generate trees that display the level of sequence similarity within each TF family (Additional data file 7). As expected, the known redundant TFs indicated above are found on adjacent branches in these trees.   (7) 26 (7) 11 (5) 257 (226) 232 (223) 293 (46) 108 (53) 3 (1) instance, although for most worm TFs the binding site is completely unknown [40], consensus DNA binding sequences are available for many human TFs and are collected in the Transfac database [41]. DNA binding domains evolve slower than other protein sequences [42] and, as a consequence, orthologous TFs recognize similar DNA sequences [43]. Therefore, DNA binding specificities of human TFs may be helpful to predict the DNA binding specificities of orthologous C. elegans TFs and vice versa. In the future, orthology of TFs will be invaluable in the study of the evolution of transcription regulatory networks.

wTF2.0: a tool for the creation of TF-ORF resources
wTF2.0 provides a starting point for the creation of physical clone resources that can be used to systematically map transcription regulatory networks. TF-ORFs can be obtained from the ORFeome resource and efficiently subcloned by a Gateway cloning reaction into various different Destination vectors [44,45] (Figure 5). To date, the C. elegans ORFeome consists of approximately 13,000 full-length ORFs, which is approximately 66% of all predicted ORFs. We searched worfdb, the ORFeome database [46], and found 652 predicted TF-encoding ORFs (70%). These TF-ORFs can be used to map transcription regulatory networks in different ways. First, they can be cloned into yeast one-hybrid prey vectors to detect physical interactions with their target genes [26]. For instance, TF-ORFs have been pooled to create a TF minilibrary that can be used in high-throughput yeast one-hybrid assays [26]. Second, they can be transferred to yeast twohybrid or 'TAG' vectors for the identification of protein-protein interactions [47,48]. This will be important to further identify functional TF complexes and to understand how TF function is regulated. In addition, TAG vectors may be useful to create transgenic worm strains that can be used in chromatin-immunoprecipitation experiments to identify TF target genes in vivo. Finally, TF-ORFs can be subcloned into an RNA interference (RNAi) vector for the analysis of loss-offunction phenotypes or for the identification of genetic inter-Protein-protein interaction network of worm TFs Figure 4 Protein-protein interaction network of worm TFs. Blue rectangles indicate homodimers. Different colors identify different TF families as indicated.

wTF2.0: a tool for the creation of TF gene promoter resources
To date, the Promoterome [11] contains approximately 6,500 promoters (33%), including 279 (30%) TF-promoters. TFpromoters can be fused to green fluorescent protein (GFP) in two configurations. TF-promoters can be fused directly to GFP in what are referred to as 'transcriptional fusions' and the resulting promoter::GFP constructs can be used to create transgenic C. elegans strains in which promoter activity can be examined by light microscopy [52]. Such lines can also be used to examine the effects on GFP expression as the result of a knockdown in regulatory TF levels by RNAi [53]. Alternatively, TF-promoters and corresponding TF-ORFs can be cloned together with GFP by multisite Gateway cloning [54] to create 'translational fusions' with GFP. The resulting promoter::ORF::GFP constructs are used to create transgenic lines in which both TF-promoter activity and TF subcellular localization can be examined [11]. Finally, TF-promoters can be cloned into yeast one-reporter vectors to identify other TFs that can physically associate with these promoters and that may contribute to TF promoter activity [26]. Such interactions are important to delineate regulatory cascades, important building blocks in transcription regulatory networks [3].

Conclusions
We have compiled wTF2.0, a comprehensive compendium of putative C. elegans TFs, using both computational queries and manual curation. Combining wTF2.0 with different physical TF clone resources provides the first step toward the systematic dissection of C. elegans transcription regulatory networks.

Prediction of C. elegans TF-encoding genes
WormPep 140 (WS140) (22,420 proteins, 19,735 genes) was searched. Proteins that have no apparent function in transcription regulation (for example, proteins involved in DNA repair and replication, chromatin remodeling, kinases) were removed. In addition, we removed general TFs (Additional data file 2). To identify TFs missed by the GO search, we searched WormPep 140 using individual DNA binding domains (Table 1). Next, we computationally (using SMART, Pfam or InterPro) or manually inspected each protein sequence for the presence of a DNA binding domain (Additional data file 2). For C2H2 zinc-fingers, we only considered proteins that contain fingers with the following configuration: C-X 2-5 -C-X 9 -H-X 3-5 -H [55]. However, we did include two proteins (LIR-1 and TLP-1) that do not have a canonical C2H2 zinc-finger, because they were found multiple times in high-throughput yeast one-hybrid assays (data not shown). For AT-hook predictions, we used the definition as described [56]. The numbers of human genes encoding each TF DNA binding domain were found by searching for the appropriate InterPro domain accession number using the Ensembl Human database.

Identification of candidate redundant TFs
For each TF family with at least five members, an alignment was created using the multiple alignment mode of ClustalX v. 1.83 [57] under default settings, including the Gonnet series of protein matrices. For TF genes encoding multiple isoforms, if the DNA binding domain was identical in all isoforms, the largest isoform was used. Isoforms containing different DNA binding domains were included separately. Unrooted trees were then generated from these alignments using ClustalX v. 1.83 using the Neighbor-Joining method, and visualized using the phylogram output of TREEVIEW PPC v1.6.6 [58] (Additional data file 7). Although the analysis was not of sufficient depth for these trees to represent real evolutionary relationships amongst the deeper branches, these trees do accurately reflect close relationships between C. elegans TFs, with candidate redundant genes occurring on adjacent short branches.

TF splice variants and alternative TF promoters
To identify TF splice variants, the coding sequences of each TF were retrieved from WS140 using the batch gene tool. Each splice variant was then manually examined for the presence of a DNA-binding domain and/or an alternative promoter.

TF dimers
TF dimers were obtained from worm interactome version 5 (WI5) [32]. TF-TF interactions were modeled into a protein interaction network using the Cytoscape software package [59]. In this network, nodes correspond to interactors and wTF2.0 can be used to create clone resources that can be used to study the transcription regulatory networks controlling metazoan gene expression Figure 5 wTF2.0 can be used to create clone resources that can be used to study the transcription regulatory networks controlling metazoan gene expression. TAG, epitope or purification tag; Y1H, yeast one-hybrid; Y2H, yeast two-hybrid.
wTF2.0 Transcription regulatory networks edges (that is, links between nodes) represent protein-protein interactions.

TF orthologs
For each TF, the best human blastp hit was retrieved from WormBase. The Ensembl ID of each retrieved human protein was then used to extract the best C. elegans blastp hit and the corresponding percentage protein sequence identity using the data-mining tool BioMart. For the 44 TFs that did not yield a hit individual blast searches were performed. For 26 C. elegans TFs still no human homolog could be identified.
Reciprocal best blast hits were considered putative orthologs and are highlighted in bold in Additional data file 8.

Additional data files
The following additional data are available with the online version of this paper. Additional file 1 is a