Now that numerous high-quality complete genome sequences are available, many efforts are focusing on the “second genomic code”, namely the code that determines how the precise temporal and spatial expression of each gene in the genome is achieved. In this regard, the elucidation of transcription regulatory networks that describe combined transcriptional circuits for an organism of interest has become valuable to our understanding of gene expression at a systems level. Such networks describe physical and regulatory interactions between transcription factors (TFs) and the target genes they regulate under different developmental, physiological, or pathological conditions. The mapping of high-quality transcription regulatory networks depends not only on the accuracy of the experimental or computational method chosen, but also relies on the quality of TF predictions. Moreover, the total repertoire of TFs is not only determined by the protein-coding capacity of the genome, but also by different protein properties, including dimerization, co-factor interactions and post-translational modifications. Here, we discuss the factors that influence TF functionality and, hence, the functionality of the networks in which they operate.

There are two classes of TFs: basal TFs that are involved in transcription of most, if not all, genes, and regulatory TFs that control only subsets of genes.20 For the understanding of differential gene expression at a systems level, we only consider regulatory TFs (hereafter referred to as TFs). TFs interact with their target genes by binding specific cis-regulatory gene elements through a sequence-specific DNA binding domain. Different DNA binding domains are used to group TFs into TF families. Examples of DNA binding domains include basic region helix–loop–helix domains (bHLH), homeodomains and various types of zinc fingers. Computational tools have been developed both to define consensus DNA binding domains and to predict additional TFs of that family encoded by a genome of interest.21,22 We found that, although such computational tools are powerful, they do incorporate false predictions and miss many known TFs.23 For instance, by using a combination of computational tools and extensive manual curation we predicted a high-quality compendium of 934 TFs in the nematode Caenorhabditis elegans, which extended purely computational predictions by ~50%.23 However, even though this compendium is more complete than previous collections, it is not yet comprehensive as algorithms and experimental assays continue to improve and, therefore, additional TFs continue to be discovered.24

Over the past decades, many different sequence-specific DNA binding domains have been uncovered. However, we propose that it is unlikely that all DNA binding domains are known. This is because, by applying yeast one-hybrid assays to only 112 C. elegans gene promoters, we have already discovered 11 C. elegans proteins that robustly interact with their target promoters in yeast, but that do not possess a known DNA binding domain. By using chromatin immunoprecipitation assays in yeast, we confirmed that these interactions are direct for nine of these proteins.10,11 We do not know yet whether these proteins directly bind to DNA or if they are recruited to their target promoters by interacting with other DNA binding proteins. Future structure–function analysis will provide insight into the mechanism of action of these novel putative TFs. Importantly, the cataloging of the DNA binding domain(s) of these proteins may enable the identification of additional proteins with similar domains in C. elegans, and perhaps in other organisms as well.

In metazoans, many gene transcripts, including those encoding TFs, are alternatively spliced, which often leads to multiple variants of a protein. Interestingly, it has been found that TF-encoding genes in mice undergo alternative splicing more frequently than other genes.25 Alternative splicing may lead to TFs with different functions. For instance, DNA binding domains or transcription regulatory domains may be included or excluded from the TF variant. At least 144 C. elegans TFs undergo alternative splicing, resulting in 379 different proteins, 30 of which lack a DNA-binding domain.23 The latter may function as regulators of TF function, for instance by titrating interaction partners of the corresponding TFs that do possess a DNA binding domain. Several C. elegans TFs contain more than one DNA binding domain and alternative splicing can affect which domains are present in the different protein products. For instance, several DAF-16 variants are generated as a result of alternative splicing, and each variant carries a unique combination of domains26 (Fig. 2(A)). DAF-16 is a critical regulator of various physiological processes including fat storage, aging, and the formation of dauers (an alternative larval stage of development, resistant to many forms of stress). It contains two potential forkhead DNA binding domains and is known to bind or regulate numerous target genes.27 It is tempting to speculate that each forkhead domain is responsible for the interaction with a distinct set of target genes, each of which may be involved in a particular biological process. A human example involves the homeodomain proteins hepatocyte nuclear factor HNF1 and vHNF1. There are three HNF1 splice variants, each of which encodes a protein with varying transcription activation properties. vHNF1 is also differentially spliced and one of the resulting protein products, vHNF1-C, functions as a trans-dominant repressor of all three HNF1 variants.28 TF variants may be expressed in different cell types or under particular conditions, leading to variable outputs of the transcriptional circuits in which they function. For instance, the HNF isoforms are differentially expressed in the human digestive tract, liver and kidney, where they may either regulate distinct target genes or, alternatively, the same target genes but at different levels. Genome-scale analyses of alternative splicing, for instance using whole genome or exon junction tiling arrays,29 will greatly facilitate the accurate identification of all TF variants that are produced in each cell and tissue type, and in various model organisms.

Several TFs bind DNA as obligatory dimers, including members from the basic region leucine zipper (bZIP), bHLH and nuclear hormone receptor (NHR) families.30–32 Dimerization should be taken into account when considering the total complement of functional TFs because, if a particular TF only functions when it dimerizes with another TF, the dimer should be considered a single functional unit. Dimerization can affect the total number of functional TFs in different ways (Fig. 2(B)). In one model, each TF from a family can dimerize with itself and any other member of that family. If this would be the case, the number of functional TFs would be dramatically greater than the number of individually predicted TFs. For instance, dimerization between all 274 C. elegans NHRs23 would result in a total of 37675 NHR dimers. In another model, each TF dimerizes exclusively with one other TF of the same family. If this were the case for the C. elegans NHRs, this would result in 137 functional TF complexes, which would reduce the total number of functional TFs. In a third, intermediate model, one or more TFs from a family could serve as central dimerization partners that can interact with multiple members of the relevant TF family. Although no comprehensive data regarding TF dimerization is as yet available, our preliminary data indicate that this third model is likely most relevant (C. A. G. and A. J. M. W., in preparation). Systematic protein–protein interaction mapping efforts will be required to identify all functional TF dimers. High-throughput yeast two-hybrid assays33 are particularly well suited for this task as they identify binary interactions. For instance, we found by high-throughput yeast two-hybrid assays that NHR-49 serves as a dimerization hub, that can interact with at least 15 other NHRs.24 In the future, it will be important not only to identify all TF dimers, but to also determine the DNA-binding specificities of each dimer and where and when dimerization partners are co-expressed.

Many proteins, including TFs, are post-translationally modified under different conditions and by different modifiers. Several post-translational modifications of TFs have been reported, including phosphorylation, hydroxylation, acetylation, ubiquitination and sumoylation (Fig. 2(C)).34 Such modifications often result from the activation of signal transduction pathways in response to environmental stimuli or developmental cues. Post-translational modifications can affect the regulatory activity of a TF, as well as its localization or stability. For instance, estrogen receptor β (ERβ), normally activated by ligand, can be phosphorylated by MAP kinase (MAPK) which leads to recruitment of the steroid receptor coactivator-1 (SRC-1) and ligand-independent activation of target genes.35 ERα can be acetylated by p300 at conserved lysine residues resulting in enhanced DNA-binding activity and, perhaps, ligand-dependent transcriptional activation.36 For most TFs it is not clear which modified forms exist and how these forms function to regulate gene expression. For a thorough understanding of gene regulatory networks, it will be important to determine which modification each TF is subjected to, under which circumstances, and how these modifications affect TF functionality.

Regulatory TFs often activate or repress transcription, either by recruiting the RNA polymerase II machinery, or by preventing its access to the transcription start site. While many TFs interact directly with general TFs or components of RNA polymerase II, others function by interacting with intermediate proteins called co-factors (Fig. 2(E)).44–46 Depending on environmental or developmental circumstances, the same TF can interact with different coactivators or corepressors, illustrating the versatility of TFs in carrying out opposing regulatory effects in different contexts. RAR, for instance, can recruit the NCoR/SMRT corepressor complexes thereby repressing transcriptional activity of its target genes. Upon binding its ligand retinoic acid, however, RAR changes conformation and adopts a form capable of recruiting coactivator complexes and subsequently activates transcription.45 Interestingly, cofactor interactions can also affect DNA binding specificity,47 implying that different TF-cofactor pairs may interact with different sets of target genes. Future large-scale genomic and proteomic experiments are needed to identify the full spectrum of ligands and co-factors each TF in an organism can interact with and to unravel how these interactions affect the biochemical and biological function of each TF.

Although complete genome sequences have provided a great first step toward the comprehensive identification of the compendium of TFs that function in an organism of interest, we are far from having a complete picture of all the protein variants that may exist for each predicted TF. As we discussed here, there are numerous factors that affect the functional states of TFs throughout development, homeostasis, and in disease. Since the gene count is strikingly similar between organisms of widely different complexity, a larger number of TF permutations may contribute to more intricate regulatory networks in higher eukaryotes such as humans. In the future, different TF forms need to be incorporated as individual nodes in transcription regulatory networks to facilitate network modeling and hypothesis derivation (Fig. 3).