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 available 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.

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.

Table 1 presents wTF2.0 grouped into TF families. Interestingly, 23 TFs contain DNA binding domains from different families (Additional data file 3). 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.

wTF2.0 is a starting point to predict the actual number of functional TF complexes. For instance, the number of active TFs is likely greater than 934 because many TF genes encode multiple proteins as a result of alternative transcripts. In addition, many TFs function as heterodimers, with subunits associating in different combinations. To date, 144 of the 934 predicted TF genes (15.4%) are known to undergo alternative splicing (Additional data file 5). On average, each spliced TF gene results in 3 different transcripts and the number of splice variants per gene ranges from 2 to 13. Some alternative transcripts do not result in the expression of a different TF protein variant. In total, 379 alternative TF protein variants are expressed from 144 genes, and the number of variant proteins per TF gene is between 2 and 10. Interestingly, 30 TF variants, corresponding to 25 TF genes, no longer contain a DNA binding domain. Rather than binding DNA and regulating target gene expression directly, these proteins may have regulatory functions to control TF activity. Taken together, alternative splicing yields 205 additional putative DNA binding TFs, bringing the total number of predicted C. elegans TFs to 1,139. Interestingly, Taneri and colleagues 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.

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 minimum 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-hybrid-based 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 protein-protein 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.

Next, we identified putative human orthologs for each worm TF, based on reciprocal best BLAST hits 38 (Additional data file 8). First, we identified members of each TF family in humans. Subsequently, we determined the number of ortholog pairs per TF family (Table 1). We found that some TF families are expanded in humans. For example, the CP2, C2HC and helix-turn-helix (HTH) TF families are represented by only one or two proteins in the worm each having a human ortholog. However, the human families are expanded six-fold. Conversely, other TF families are expanded in the worm, compared to human. As reported previously 39, the NHR family is expanded in worms and contains 274 predicted members (versus 43 in humans). Interestingly, the MADF family, which is composed of nine proteins in the worm, is not found in humans.

Putative TF orthologs may comprise a valuable tool to annotate TF function in either worm or human systems. For 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 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 4445 (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 mini-library that can be used in high-throughput yeast one-hybrid assays 26. Second, they can be transferred to yeast two-hybrid or 'TAG' vectors for the identification of protein-protein interactions 4748. 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-of-function phenotypes or for the identification of genetic interactions 495051. Phenotypic analyses of TFs will be important for the analysis and interpretation of transcription regulatory networks.

To date, the Promoterome 11 contains approximately 6,500 promoters (33%), including 279 (30%) TF-promoters. TF-promoters 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.

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-X₉-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.

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.

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 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 edges (that is, links between nodes) represent protein-protein interactions.

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.