Access to the DBD database is via http://transcriptionfactor.org, where all data is available for viewing and immediate download. The community can browse predictions for over 700 species (from Arabidopsis thaliana to Zymomonas mobilis) or DBD family (including helix–turn–helix, zinc-fingers, homeobox and many others); search predictions by sequence identifier or domain family; receive classifications for submitted protein sequences, and download our domain assignments, as well as our manually curated list of DBDs.

The prediction method in the DBD database (10) uses hidden Markov models (HMMs) to identify domains in proteins from two databases: SUPERFAMILY (11) and Pfam (12). From DBD release 2.0 onwards, updated annotation resulted in 303 HMMs from SUPERFAMILY and 145 from Pfam compared to a total of 251 HMMs in the first version of DBD. The HMMs from SUPERFAMILY represent 37 superfamilies and 87 families according to the definitions in the SCOP database (13). This includes 98 new models representing 37 sequence-specific DBD families. This resulted in an increase in additional TF predictions of 4.7%, for the 150 organisms in the original version of DBD.

The pipeline used to predict TFs begins with a domain annotation of all proteins from completely sequenced genomes with all HMMs from the SUPERFAMILY and Pfam databases (Supplementary Figure 1). A protein is classified as a TF if it has a significant match to a model we annotated as being a DBD, with the significance thresholds for HMM matches taken from the Pfam and SUPERFAMILY databases. This results in an estimated 1–5% of false-positive annotations. The TF predictions are limited to the families in our annotated collection, which means that the coverage is about two-thirds of known TFs. At the same time, up to an additional 50% of proteins are predicted as TFs that have annotations such as ‘hypothetical protein’, particularly in metazoan genomes. For details of benchmarking, please refer to (10). The prediction method is general and applicable to any proteome or sequence set. In fact, the database has grown to encompass TF repertoires of over 700 publicly available genomes. Predictions for newly sequenced genomes are continuously added to the database.

The current DBD database contains information on over 200 000 predicted TFs. These TFs are distributed across the tree of life. It is not surprising that, we find a greater number of TFs in larger genomes. To investigate the relationship between TF abundance and proteome size in different lineages we graph these variables on a log–log plot as in Kummerfeld and Teichmann (10) (Supplementary Figure 2 in this paper). To illustrate the difference between the eukaryotic and prokaryotic superkingdoms we separately perform a model fitting for these lineages. From the linear relationship on the log–log scale a power law can be inferred. This power law could be due to the underlying distribution of DBDs. A small number of DBDs (such as helix–turn–helix and zinc-finger families) occur in the majority of TFs. Whereas most DBDs occur in only a small number of TFs. In agreement with van Nimwegen (14) and Ranea et al. (15), we find a higher proportion of TFs are required to regulate larger proteomes. We also find the TF abundance in archaea and bacteria expands more rapidly than in eukaryotes. Thus, in general, the same number of TFs regulate fewer prokaryotic genes than eukaryotic genes. The higher degree of combinatorial control, where gene expression is regulated by not just one but by a group of TFs, may also contribute to the lower eukaryotic TF requirements. Different combinations of TFs mean the number of gene regulation modes can increase with a reduced increase in TFs. Bacteria and archaea obey the same power law in terms of number of TFs and number of proteins. This is in accordance with their shared repertoire of DBD families, which we will return to below.

We illustrate the TFs containing a specified DBD family in a particular organism in Figure 1, where a hypothetical researcher is interested in the Homeobox TFs. These TFs are known to regulate vertebrate limb formation amongst other processes (16). Figure 1a depicts the search for TFs in Homo sapiens containing the homoeobox domain. A subset of the results of this search are shown in Figure 1b. By selecting the HOXA9 TF from this result set, the researcher can examine one of the new pages containing detailed information on each TF (Figure 1c). The detailed pages include the sequence of the TF, links to external databases containing further information on the protein, domain assignment regions and an indication of the quality of the domain assignment in the form of an Evalue. Links to predicted TFs with similar domain combinations are also provided on these pages. An example of predicted TFs with similar Pfam architectures to the HOXA9 TF (i.e. an N-terminal Hox9 activation region and a C-terminal Homoeobox domain) is shown in Figure 1d.

Using the data on DBD families in different organisms, we compare the occurrence of DBDs (from the Pfam project) across the tree of life. The heatmap in Figure 2 demonstrates the lineage-specific DBD expansions and contractions. The list of species and DBD lists are included in Supplementary Tables 1 and 2. We found the number of occurrences of each DBD in each organism, and then normalized this number by the proteome size of that organism. In order to represent both contractions and expansions, we calculated a Z-score for each of the normalized DBD occurrence values. The Z-score is calculated from the distribution of normalized DBD occurrence across genomes for a particular DBD family, and has a mean of zero and a standard deviation of one. It is negative when the normalized DBD occurrence is below the mean, and positive when above the mean. In Figure 2, DBD expansions (positive Z-scores) are represented using red, and contractions (negative Z-scores) using green.

Different sets of DBDs expand in different lineages. There is a clear separation between the DBD occurrence pattern in eukaryotes (in the top section of the heatmap) and prokaryotes. The DBD occurrence in prokaryotes is relatively diverse. For instance, there is a significant overlap between the DBD repertories of the actinobacteria, proteobacteria and firmicutes. This is almost certainly due to the ubiquitous horizontal gene transfer between prokaryotes. The DBD expansion pattern in archaea is similar to that in bacteria, despite sharing conserved basal transcriptional machinery with eukaryotes rather than with bacteria. The majority of these prokaryotic DBDs have the helix–turn–helix as part of their structure (17).

The eukaryote-specific DBD expansions have considerably greater variety than the prokaryotic expansions. An increased DBD kingdom-specificity is found in the eukaryotes. The metazoan, fungal and plant kingdoms are clearly distinguishable (Figure 2a). The fungal and metazoan kingdoms share more DBDs than the plant and metazoan kingdoms, which reflects their closer phylogenetic relationship (18). The metazoa, in the top right section of Figure 2a, have the largest kingdom-specific DBD repertoire. This is most likely due to the regulatory overhead of metazoan complexity in terms of cell types.

We have eliminated several eukaryotic genomes (Xenopus tropicalis, Apis melifera and Populus trichcarpa) from our analysis of DBD occurrence due to the presence of uncharacteristically high numbers of bacterial DBDs. This was a known problem in the X. tropicalis (frog) genome (22). The use of lineage-specific information on the occurrence of DBDs is a promising method for reducing false-positive TF classifications in the eukaryotes.