Regulation of gene expression in an organism involves a complex network. DNA-binding transcription factors are an important component of this network: they respond to changes in the cellular environment by altering the gene expression of relevant genes. Due to this crucial role of transcription factors, they have been studied in many ways, including elucidation of numerous three-dimensional structures. Theoretical analyses of transcription factors in Escherichia coli have focused on their sequence families and sequence motifs (1,2). In addition, recent research has elucidated the design principles of the transcriptional regulation network (3–5), including the motifs that recur in the network and their functions.

Our approach is based on the determination of the homology between the domains and protein families of transcription factors and regulated genes, and proteins of known three-dimensional structure. This provides a powerful tool, which goes far beyond sequence comparison methods alone, for finding the domain architecture and evolutionary relationships of transcription factors. Using this method, we can identify uncharacterised E.coli proteins that contain DNA-binding domains (DBDs) and identify what is likely to be the large majority of E.coli transcription factors.

The homologies between the transcription factors and proteins of known three-dimensional structure yield the domain compositions of the transcription factors. This allows us to quantify the features of this repertoire of transcription factors for E.coli: we find that three-quarters of the transcription factors are two-domain proteins, a trend noted previously by Morett and Segovia (6) and Aravind and Koonin (7), and we establish that half of them bind small molecules, a phenomenon first discovered by Jacob and Monod (8).

Based on the domain architectures of the known and predicted transcription factors, we can trace the duplications and recombinations that have produced these proteins in E.coli in a more general and extensive manner than has been previously possible (1,7,9). This analysis of domain architecture shows that three-quarters of the transcription factors have arisen by gene duplication.

For the subset of experimentally studied transcription factors, there is information available about the genes they regulate. This allows us to classify these transcription factors in terms of their functions and the numbers of transcription factors and other genes they regulate. We have collated the complete set of transcription factors regulating other transcription factors into a single figure coloured according to the evolutionary family of the DBD, which provides an overview of the central network of gene regulation in E.coli at a glance. In order to gain insight into evolution of the entire regulatory network, we have looked for instances of duplications of regulated genes together with their regulatory regions, as well as for duplications of transcription factors together with regulated genes.

In SCOP, evolutionary relationships of domains of known structure are inferred through a combination of clues from sequence, structure and function. Since protein structure is more conserved than sequence in evolution, the structural domain and family definitions in SCOP are much more accurate and extensive than could be achieved by sequence comparisons alone. We refer to the SCOP superfamilies as protein families throughout this work.

The SUPERFAMILY database of structural assignments based on SCOP domains uses hidden Markov models (14), which are probably the most sensitive automatic sequence comparison method currently available (15). The procedure used to make the hidden Markov models and scan them against complete genomes is the iterative SAM-T99 method (16).

As described in Figure 1, SUPERFAMILY assignments were retrieved for the set of 416 proteins in the E.coli genome with a DBD assignment. We filtered the set of 416 proteins with DBDs by removing proteins involved in replication/repair, transposases and restriction enzymes according to the functional annotations in GenProtEC (17) and the COGs database (18). We did not include the four σ factors that have structures homologous to the σ⁷⁰ subunit fragment of RNA polymerase (rpoD, rpoS, rpoH, fliA) nor rpoN, which does not have any assigned structure. This resulted in a final set of 271 transcription factors with DBD assignments from the SUPERFAMILY database. In addition to the structural assignments from SUPERFAMILY, 46 of the transcription factors had domain assignments from nine families from the Pfam database (19) of hidden Markov models.

Figure 1 Flow chart of the method used for identification of transcription factors. In addition to our set of 271 transcription factors, there are eight transcription factors without a DBD assignment that have known regulatory information.

For 121 of the 271 transcription factors, there is experimental information about the genes they regulate in the RegulonDB database (20) and in Shen-Orr et al. (4) as well as two references about FIS not in either data set (21,22). For eight of the transcription factors with known regulatory information there were no homologues of known structure detected in the SUPERFAMILY database. Given that 113 out of 121 transcription factors with known regulatory information were assigned a DBD, it is likely that the 271 transcription factors assigned a DBD represent the large majority of all E.coli transcription factors. Other calculations have given an upper limit of 400 (3) and 350 (1) transcription factors. Thus our analysis encompasses a sizeable fraction of the entire repertoire of E.coli transcription factors and the conclusions we draw based on 271 proteins are likely to hold for the whole set.

Figure 2 (A) (Opposite and above) The three-dimensional structures of the 11 DBD families seen in the 271 identified transcription factors in E.coli. The figure highlights the fact that even though the helix–turn–helix motif occurs in all families (more ...)

Given that the set of 271 E.coli transcription factors have only 11 different DBDs, it is of interest to analyse the distribution of DBDs in these proteins. As shown in Table 1, the sizes of the DBD families vary from 123 members of the winged helix family to a single member in the Trp repressor and nucleic acid binding domain families. The homeodomain-like family is the second largest family with 52 domains; there are two other medium-sized families and the remaining DBD families are small, following a power law type distribution of family size similar to that observed in complete genomes (23,24).

Table 1. Information about the domain architectures of each DBD family

Three-quarters of transcription factors are two-domain proteins. The DBDs generally occur in combination with other domains: there were only 25 single-domain proteins (~10%), but 202 two-domain proteins (~75%), 33 three-domain proteins (~12%) and nine four-domain proteins (~3%). All proteins contain a single DBD except for an uncharacterised protein that contains two copies of winged helix domains and 20 proteins that contain two adjacent homeodomains. There are two separate crystal structures of E.coli proteins with two adjacent homeodomains; in one structure only one of the homeodomains interacts specifically with the major groove and in one of them both domains do so (25,26).

Instead of a second copy of the DBD, the other domain more frequently has a different function, such as a small molecule-binding or enzymatic domain. The set of non-DBDs comes from 46 different families. These 46 families can be classified into five broad functional categories. There are 12 families of enzyme domains, of which at least three are certainly catalytically active: the signal peptidase domain in LexA, the methyl-DNA protein methyltransferase as in Ada and the P-loop nucleotide triphosphate hydrolase in DnaA. For the other nine enzyme domains, it is unclear whether they are catalytically active, as there are known examples where an enzymatic domain has lost its ability to catalyse a reaction and just serves as a small molecule-binding domain (27). There are 18 families of small molecule-binding domains, five protein interaction domain families and 10 domains of unknown function. Finally, there are CheY-like response regulator receiver domains that are phosphorylated by kinases in two-component signal transduction systems. These different categories are indicated by the shapes of the domains in Figure 2B. The large DBD families have partner domains from all of the functional categories and the partner domains can be positioned N- or C-terminal to the DBD, as shown in Table 1.

Overall, small molecule-binding domains are the most frequent type of partner domain, occurring in 44% of the transcription factors. In addition, the Trp and Met repressor DBDs bind small molecules with the same domain that binds DNA. This suggests that almost half of the transcription factors in E.coli are directly regulated by the presence or absence of small molecules, as previously noted (27).

The CheY-like response regulator receiver domain occurs in ~10% of the proteins. Protein interaction domains that either interact with RNA polymerase subunits or are involved in dimerization occur in ~7% of the proteins. Enzymatic domains occur in ~5% of the transcription factors. One or two DBDs occur in isolation in ~12% of the proteins. In the remaining the cases, the DBD occurs in combination with a domain of unknown function, or a region for which no known domain assignment can be made.

This distribution of partner domains suggests that only a small fraction of transcription factors are regulated exclusively at the transcriptional level and not by a small molecule or through a sensor protein. The major types of domain combinations are a DBD with a small molecule-binding domain or a Che-Y like response regulator receiver domain with the C-terminal effector domain (25 proteins). There are 120 proteins with 27 distinct combinations of domain families of the DBD with small molecule-binding domain type: the two main domain architectures are winged helix with periplasmic binding protein-like II (43 proteins) and the λ repressor-like with the periplasmic binding protein-like I (14 proteins).

Three-quarters of the E.coli transcription factors have arisen by gene duplication. From Table 1 and the schematic representation of the domain architectures of the transcription factors in Figure 2B, it is obvious that these proteins have evolved by extensive recombination of domains. However, proteins with the same sequential arrangement of domains are likely to be direct duplicates of each other, as discussed in Apic et al. (28) and Bashton and Chothia (29). Therefore, we have also looked at whole proteins rather than individual domains and have grouped them into protein families maintaining the same domain architecture. In total, there are the 74 distinct domain architectures shown in Figure 2B, which have duplicated to give rise to 271 transcription factors. Thus 73% of these transcription factors have arisen as a consequence of complete gene duplication. The protein families maintaining the same domain architecture can contain members with different regulatory activities: there are several domain architectures that are found in both activators and dual regulators or repressors and dual regulators (30).

In the FIS-like DBD family, it is obvious that a two-domain fragment has duplicated with subsequent recombinations with one or two additional domains, so that the DBD forms an evolutionary module with a P-loop-containing nucleotide triphosphate hydrolase. In these proteins, the P-loop domain interacts with the σ⁵⁴ subunit of RNA polymerase. There are three different domain architectures that have a GAF domain N-terminal to these two domains and three other domain architectures in which a PYP-like sensor domain is N-terminal to the two-domain module. This example illustrates how a module of two domains acts as an evolutionary unit that is elaborated to different three-domain modules.

In contrast, there are two examples where a pair of domains is inverted rather than retaining the N- to C-terminal order. Domains of the winged helix family occur both N- and C-terminal to periplasmic binding protein II domains. The C-terminal effector domain of the bipartite response regulator occurs N-terminal to TPR repeat domains in two different architectures and C-terminal in one domain architecture.

It is worth mentioning that the winged helix DBD almost always occurs at the N-terminus, as shown in Table 1. The only exception is the cAMP-binding domain-like family, which occurs N-terminal to the winged helix DBD. The four proteins with this domain architecture are CRP and FNR, which are both global regulators controlling a large number of genes, and two hypothetical proteins.

Table 2. Regulated genes and functional information for the global regulators (for a complete list please refer to www.mrc-lmb.cam.ac.uk/genomes/madanm/ec_tf/)

Regulatory cascades: the central part of the transcriptional network. Individually, these 121 transcription factors regulate from 1 to 197 genes and, all together, there are 1302 genes and 303 operons in the regulatory network (transcription factors and regulated genes). For 38 of the transcription factors, some of the regulated genes are themselves transcription factors. There are 34 autoregulatory transcription factors, as listed in Rosenfeld et al. (31). To investigate the regulation of transcription factors, we integrated the information available from RegulonDB (20) and Shen-Orr et al. (4), as well as Falconi et al. (21) and Gonzalez-Gil et al. (22), to produce the diagram in Figure 3. This figure shows the network of transcription factors currently known to regulate each other in E.coli. CRP controls 18 different transcription factors apart from itself. Two other transcription factors, FNR and ArcA, regulate four transcription factors and FIS and IHF (himA and himD) regulate three transcription factors. CRP is a global sensor of food levels in the environment. FNR and ArcA are involved in sensing the redox status of the cell to regulate genes involved in respiration under aerobic and anaerobic conditions. FIS and IHF-like are general enhancers, which frequently act together with other transcription factors to regulate genes. Thus the transcription factors involved in respiration and growth are those which regulate the most transcription factors.

Figure 3 The transcription factor regulatory network in E.coli. When more than one transcription factor regulates a gene, the order of their binding sites is as given in the figure. An arrowhead is used to indicate positive regulation when the position of the (more ...)

In fact, there are few long cascades of transcription factors that regulate each other in the E.coli gene regulatory network, as noted previously (3,4). In our current dataset as illustrated in Figure 3, there are 23 two-level cascades, 32 three-level cascades and six four-level cascades. Thus even in the simple prokaryote E.coli, the transcriptional regulation network is a complex combination of multi-level cascades and motifs.

Global regulators. Through regulation of other transcription factors, the transcription factors amplify their range of control over genes to encompass a set of indirectly regulated genes. Thus the total number of genes regulated by a transcription factor is the sum of the genes regulated directly and indirectly, given in the third column of Table 2. In 6 of the 10 functional categories, there are transcription factors that regulate genes both directly and indirectly, and control more than 15 genes all together. These 11 transcription factors, shown in Table 2, can be viewed as ‘global regulators’, as opposed to the remaining transcription factors, which are ‘fine tuners’. For a complete list of transcription factors, please refer to the supplementary data at www.mrc-lmb.cam.ac.uk/genomes/madanm/ec_tf/. The difference between these two types of transcription factors is clear from the graph in Figure 4: the global regulators have more directly and indirectly regulated genes than the remaining transcription factors. In a recent analysis of the E.coli network motifs by Shen-Orr et al. (4), global regulators were defined in a different way than here: as those transcription factors that controlled 10 or more operons. This gives a set of 15 global regulators, nine of which are also in our set of 11 global regulators.

Figure 4 Direct and indirect gene regulation by E.coli transcription factors. The direct number of genes regulated is represented on the x axis and the indirect number of genes on the y axis. The global regulators, which are marked on the graph regulate a large (more ...)

Our set of 11 global regulators are transcription factors involved in carbon degradation (mlc and lrp), redox status sensing (arcA, narL and FNR), ion transport regulation (fur), environmental sensors (cspA and CRP), a regulator of structural proteins (hns) and two general enhancers (IHF and FIS). Thus the global regulators are proteins that control responses to changing food levels and carbon degradation (mlc, lrp, CRP) and transcription factors that respond to changes in redox status or ion levels of the cell (arcA, narL, FNR, fur). Cold shock protein A (cspA) binds RNA and regulates translation in this manner, but it is also known to bind DNA (32). FIS is a homeostatic regulator of general superhelicity.

Eight of the 11 global transcription factors are dual regulators. mlc and fur are only repressors, and cspA is only an activator. cspA is the only global transcription factor that has a DBD of the nucleic acid-binding protein family; the other transcription factors belong to three other DBD families and have seven different domain architectures.

With the current status of experimental data, the remaining 109 transcription factors each regulate 33 or fewer genes in total. Therefore, with the current status of information about the E.coli gene regulatory network, it appears that the majority of transcription factors are ‘fine tuners’ that control a limited, specific set of genes, while a small number of transcription factors are ‘global regulators’ that control tens or hundreds of genes by direct and indirect influence.

Figure 5 Distribution of the number of transcription factors regulating a gene. Numbers in parentheses represent the number of operons.

There is evidence from other organisms that the same pair of transcription factors has adjacent binding sites in many regulatory regions in the genome. In yeast, such synergistic pairs of transcription factors were studied by Pilpel et al. (33), and Berman et al. (34) analysed clustered binding sites of five transcription factors active in the early Drosophila embryo. In the current data set, 24 pairs of transcription factors regulate between two and five operons, and four triplets of transcription factors regulate two or three operons.

Table 3 shows the distribution of transcription factors that are present at the same promoter as other transcription factors. Seven transcription factors occur at the same promoter with over 10 different transcription factors, and CRP occurs with 52 other transcription factors. This large variation of combinations suggests that not all, if any, of these transcription factors interact physically in a specific manner, such that the interactions with the DNA and the RNA polymerase are the decisive ones. This is supported by experiments such as those of Martin et al. (35).

Table 3. Numbers of transcription factors active at the same promoter

Genes with similar regulatory regions. Given that regulatory regions are composed of binding sites for one or more transcription factors as described above, we want to address the evolution of the regulatory regions by looking for evidence of duplications of regulatory regions with their regulated genes. We define such duplications as operons of homologous genes which are regulated by the same transcription factor(s). Homologous genes are those whose protein products have the same domain architecture according to SUPERFAMILY domain assignments.

Twenty such cases are shown in Table 4. A rough indication of the duplication rate of operons and their regulatory regions can be obtained by dividing these 20 duplicates by the 303 operons with genes with structural assignments in our set. This is a 7% duplication rate, compared to the three-quarters of transcription factors that have evolved by gene duplication. Since a transcription factor regulates genes that are functionally and not evolutionarily related, one would not expect a particularly high level of duplication of regulatory regions together with downstream operons. However, the results presented here and in Rajewsky et al. (36) show that duplication does contribute to the evolution of the regulatory network.

Table 4. Homologous genes regulated by the same transcription factor

Two examples of individual transcription factors regulating homologous genes are given in Figure 6A. Figure 6B shows cases of pairs of transcription factors regulating homologous genes. In the case of hupA and hupB shown in Figure 6B, both have one binding site for CRP and four for FIS. However, the four FIS-binding sites of hupA are all upstream of the transcription start and FIS is an activator, while for hupB, one of the sites is downstream and FIS is a repressor. In fact, the numbers and positions of transcription factor-binding sites are actually very different in nine of the 20 cases of homologous genes regulated by the same transcription factor(s).

Figure 6 (Opposite and above) Duplication of regulatory modules in the network. (A) Two sets of homologous genes regulated by the same transcription factor. In the first example, three genes forming part of the same operon are homologous to three genes in separate (more ...)

Homologous transcription factor-regulated gene modules. In the previous section, we considered similar regulatory regions and possible duplications of genes together with their regulatory regions. We can extend this to combinations of a transcription factor and regulated gene that are both homologous to another transcription factor and its regulated gene. This would provide evidence for growth of the regulatory network through duplication of sections of the chromosome that include a transcription factor and its regulated gene.

The sets of transcription factors and regulated genes that have homologues are shown in Table 5. In the cases of autoregulation, the transcription factor and regulated gene are the same. There are 28 transcription factors with regulated genes that may have evolved by duplication out of 303 sets in total. This represents a duplication level of 9%, which is very small compared to gene duplication levels amongst the transcription factors or compared to domain duplication levels generally found in genomes (23).

Table 5. Homologous transcription factors and regulated genes

Two examples of module duplication are given in Figure 6C. In both of the examples, the transcription factors are located next to the regulated genes on the chromosome, suggesting that the regulatory module may have duplicated as one unit. In both cases, the arrangement of the operons is slightly different. In addition, the numbers and positions of transcription factor-binding sites is different, though this is not shown in the figure. Based on the data we have here, duplication of both regulatory regions and genes, or of transcription factors together with regulated genes, plays a minor role in the evolution of the gene regulatory network in E.coli.

By using domains assigned to E.coli transcription factors through homology to proteins of known three-dimensional structure, we can accurately determine the domain architectures and evolutionary relationships of this repertoire of proteins for the first time. Previous analyses of large numbers of predicted E.coli transcription factors have focused on the small helix–turn–helix motif, and largely neglected the fact that this is part of different families of DBDs and their combinations with different families of partner domains (2,9). With our approach of using structural assignments to E.coli proteins, we identified 271 proteins as transcription factors with DBDs. This set is likely to represent a large fraction of all transcription factors.

This set of 271 transcription factors identified by us have DBDs from 11 different families. About three-quarters of these transcription factors are two-domain proteins with one DBD and one control domain, which, most frequently, is a small molecule-binding domain. By grouping the transcription factors according to their domain architectures, we found that almost three-quarters of the transcription factors have evolved as a consequence of complete gene duplication.

The rate of duplication of regulatory modules is much lower: only 7% of regulated operons have homologous genes regulated by the same transcription factor and 9% of transcription factor-regulated operon modules have homologues. This suggests that the individual elements of transcription factors, regulatory regions and regulated genes mainly evolve separately.

The set of transcription factors can be classified into 10 broad functional classes, and in certain of these there are global regulators that control many genes. Eight of the 11 global regulators are in the following four categories: carbon degradation, carbon metabolism, redox sensing and control of ion transport. The global regulators amplify their effect by regulating other transcription factors, and overall 38 of the 120 transcription factors with regulatory information control other transcription factors. These transcription factors that regulate other transcription factors are collated in a single figure, including information about the DBD family of the proteins. This figure provides a summary of the central part of the transcriptional network currently known in E.coli, and reveals that there are a small number of multi-level regulatory cascades amongst the transcription factors.

We acknowledge Julio Collado-Vides and Heladia Salgado for readily providing us with information from RegulonDB, Julian Gough for help with the SUPERFAMILY structural assignments and Cyrus Chothia, Graeme Mitchison and Andrew Travers for comments on the manuscript. We are grateful to the Medical Research Council, Cambridge Commonwealth Trust and Trinity College, Cambridge, for financial support.