Bacterial signaling mechanisms are extremely diverse, ranging from simplest two-domain transcription regulators, such as AraC or LacI, to multi-component signaling cascades that regulate sporulation, flagellar biosynthesis or biofilm formation. Until recently, the term 'signal transduction' has been typically reserved for the two-component systems consisting of a sensor histidine kinase (HK) and a response regulator (RR). In keeping with this tradition, I did not include in this survey single-component transcriptional regulators, whether of AraC type 52 or much more complex NorR type 53 and considered only dedicated signaling systems that consist of more than two individual components. In addition to HKs, these included Ser/Thr protein kinases, adenylate and diguanylate cyclases and two types of predicted c-di-GMP phosphodiesterases, containing, respectively, EAL or HD-GYP domains. Other enzymatic output domains as well as DNA- or RNA-binding response regulators have not been considered here but could be added to the list in the future. Because of the previously noted parallelism between the domain architectures of intracellular signaling proteins (e.g. PAS-GGDEF-EAL) and respective response regulators (e.g. CheY-PAS-GGDEF-EAL) 3, no attempt has been made to distinguish such proteins; they were counted both in the GGDEF and EAL columns. Naturally, such proteins were counted only once to obtain the total number of signaling proteins encoded in any given genome.

The data set included complete bacterial and archaeal genomes sequenced by the end of 2004. While Archaea and Bacteria are generally considered separate domains of life in the prokaryotic world, there are indications that many signal transduction systems in archaea have been acquired from bacteria through lateral gene transfer 232. Hence, for the purposes of this study, domain Archaea was treated as just another bacterial phylum. Owing to the redundancy of the current genome list, only one representative genome per species was used in the analysis, typically the first one to be publicly released. Exceptions included two strains of Escherichia coli, K12 and O157:H7 5455, and three serovars of Salmonella enterica, Typhi, Typhimurium, and Paratyphi 565758.

The total numbers of copies of each signaling domain encoded in each given genome were estimated in iterative PSI-BLAST 59 searches, using the strict inclusion threshold expect values of 10^-7–10^-4, adjusting as necessary. Potential false-positive hits were checked at every step of PSI-BLAST using the CDD Domain viewer 60 and manually removed (unselected) from the hit list for the next iteration of PSI-BLAST. The most typical sources of the false-positive hits were as follows.

Histidine kinases consist of two separate domains, (i) a well-conserved ATPase domain of the GHKL family 6162, referred to as HATPase_c domain [Pfam:PF02518 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF02518] in the Pfam database 14, and (ii) a less-conserved phosphoacceptor (dimerization) domain, carrying the phosphorylatable His residue 763. The dimerization domains are quite diverse in their sequence and comprise the His Kinase A (phosphoacceptor) domain clan in Pfam, which unifies four individual domain families: HisKA [Pfam:PF00512 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF00512], HisKA_2 [Pfam:PF07568 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF07568], HisKA_3 [Pfam:PF07730 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF07730, and HWE_HK [Pfam:PF07536 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF02518]. Due to the great variability of the HisKA domains, the results of PSI-BLAST search are largely determined by the presence of the HATPase_c domain and often include other members of the GHKL family, such as DNA gyrase B and DNA repair protein MutL, as well as anti-sigma F factors (SpoIIAB-like Ser/Thr kinases). Due to the presence of long α-helices in the phosphoacceptor domains, they sometimes show spurious low-complexity hits.

Methyl-accepting protein (MCP) domain (PF00015) [Pfam:PF00015 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF00015] contains long α-helices, which also attract low-complexity hits. However, the extremely high conservation of the (LI)LALNAAIEAARAGExGRGFAVVAxEVR sequence pattern allows a relatively easy recognition of false-positive hits.

Ser/Thr/Tyr kinase (STYK) domain (PF00069) [Pfam:PF00069 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF00069] belongs to the Protein kinase superfamily clan in Pfam 14. Other members of this clan, such as kinases of kanamycin, streptomycin, methylthioribose, homoserine, choline, and 3-deoxy-D-manno-octulosonic acid (KDO), are often retrieved in PSI-BLAST searches. In fact, the latter enzyme, KDO kinase (product of the waaP gene, PF06293 [Pfam:PF06293 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF06293]) often gives much better BLAST scores than certain divergent Ser/Thr kinases. Most of the discrepancies between the data presented here and those in the KinG database 6465 could be attributed to those false-positive hits. The most common false-negative hits were the putative protein kinases of ABC1/AarF family (PF03109 [Pfam:PF03109 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF03109] or COG0661 [http://www.ncbi.nlm.nih.gov/COG/new/release/cow.cgi?view=1&cog=COG0661]), which are somehow involved in ubiquinone biosynthesis, most likely by regulating this pathway 66. It should be noted that although members of the ABC1 (

a

bc1

GGDEF domains (PF00990 [http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF00990] from diverse bacteria have diguanylate cyclase activity 2324 and are structurally related to the eukaryotic adenylate cyclase (AC3) domains 67. While PSI-BLAST searches of GGDEF domains rarely produced any false positive hits, many GGDEF-related domains appeared to be inactivated, some were clearly truncated. The latter ones were excluded from the total count. The most interesting example included a conserved family of proteins (COG3887 [http://www.ncbi.nlm.nih.gov/COG/new/release/cow.cgi?view=1&cog=COG3887]) comprising a fusion of a modified (likely inactivated) GGDEF domain and the DHH-family (PF01368 [Pfam:PF01368 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF01368], 68) phosphoesterase domain. Members of this family are encoded in genomes of most Firmicutes, including tiny genomes of some Mycoplasma spp., but their function remains unknown.

EAL, AC1, AC2, or AC3 domains (corresponding to the Pfam entries PF00563 [Pfam:PF00583 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF02518], PF01295 [Pfam:PF01295 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF01295], PF01928 [Pfam:PF01928 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF01928], and PF00211 [Pfam:PF00211 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF00211], respectively) did not return any false-positive hits in PSI-BLAST searches.

HD-GYP domain is a variant of the widespread HD-type phosphohydrolase (PF01966 [Pfam:PF01966 http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF01966], 69) domain that contains a C-terminal subdomain with extra conserved residues 26. Classical HD domains without the second subdomain often showed up as false-positive hits; these were filtered based on the total length of the BLAST alignment.

Whenever possible, the domain and protein counts were compared to the published data and all discrepancies were manually verified. Thus, this census has identified 92 HKs in Bradyrhizobium japonicum, 62 HKs in Mesorhizobium loti, and 48 HKs in Sinorhizobium meliloti [see Additional file 1], which was much more than 80, 47 and 40 HKs, respectively, recognized in these bacteria in a recent survey 34. A comparison of the two sets revealed that most of the proteins missing from the HK list by Hagiwara et al.34 comprise a conserved family (COG3920 [http://www.ncbi.nlm.nih.gov/COG/new/release/cow.cgi?view=1&cog=COG3920]) with an unusual HisKA_2 (PF07568 [http://www.sanger.ac.uk/cgi-bin/Pfam/getacc?PF07568]) dimerization domain, which, however, still contains a conserved His residue, confirming that these proteins are true HKs. This and other comparisons showed that, in most cases, different authors correctly identified the core sets of signaling proteins and most discrepancies could be attributed to the different ways of treating divergent, inactivated and truncated sequences. The approach adopted here was to take a middle ground, not counting clearly truncated and highly diverged sequences but keeping in the list full-length domains that might have had inactivating point mutations. For example, although Gly?Ala and Glu?Ala changes in the GGEE motif of the GGDEF domain have been shown to abrogate its diguanylate cyclase activity, sequences with such changes were still counted as diguanylate cyclases, while the truncated sequences in Methanococcus kandleri protein MK0296 [UniProt:Q8TYK1 http://www.expasy.org/uniprot/Q8TYK1], Aeropyrum pernix protein APE1864 [UniProt:Q9YAS9 http://www.expasy.org/uniprot/Q9YAS9, or in COG3887 [http://www.ncbi.nlm.nih.gov/COG/new/release/cow.cgi?view=1&cog=COG3887] proteins (see above) were not. Likewise, Archaeoglobus fulgidus encodes a family of proteins that have a typical HK domain architecture but lack the HATPase domain. Such truncated sequences were not included in the total count [see Additional file 1] but still listed (marked with asterisks) in the supporting files. Since the signaling protein count was based on the domain count, monster multidomain proteins, combining various output domains, such as the hybrid HK-STYK [UniProt:O32393 http://www.expasy.org/uniprot/O32393] described in Spirulina platensis 70 or the HK-GGDEF combination, found in Geobacter sulfurreducens protein GSU3350 [UniProt:Q747B7 http://www.expasy.org/uniprot/Q747B7], have been counted more than once.

The census of signal transduction proteins encoded in complete microbial genomes [see Additional file 1] revealed several interesting trends. It has largely confirmed previous observations 2471 that the total number of regulatory proteins encoded by each given organism genome positively correlates with the genome size (Figure 2a) and the total number of encoded proteins (Figure 2b): microbes with complex life styles generally have larger genomes and encode more sophisticated and diverse regulatory systems than parasites with their largely degraded genomes.

While small genome size (and the correspondingly low number of signaling systems) is often associated with pathogenicity, there are numerous pathogens with relatively large genomes (e.g. Bordetella parapertussis, Mycobacterium tuberculosis), as well as free-living organisms with very small genome sizes (e.g. Thermoplasma acidophilum, Aquifex aeolicus, see Figure 2a). Many free-living archaea encode very few, if any, proteins involved in signal transduction. For example, among 2977 proteins, encoded by the extreme thermoacidophile Sulfolobus solfataricus, only 9 are signaling (8 STYKs and an AC2). A similar picture has been reported in marine cyanobacteria 72 and is seen in the recently sequenced genome of the ruminal bacterium Mannheimia succiniproducens, which encodes just 5 HKs, an AC, and a STYK [see Additional file 1]. Apparently, the constant and nutrient-rich ruminal environment does not require much signal transduction. These data indicate that organisms inhabiting stable environments can get away with relatively simple signal transduction systems. In contrast, organisms that survive in diverse ecological niches, including facultative pathogens, such as the spirochetes Leptospira interrogans and Treponema denticola, typically encode complex sensory systems. Of course, sophisticated bacteria can also be found in simple and stable environments: Wolinella succinogenes, another ruminal inhabitant 73, encodes many more signal transduction proteins than other bacteria with similar genome sizes (Table 1, see below).

The total number of signaling proteins encoded in a given genome (or, rather, the fraction of such proteins among all encoded in the genome) can be used as a measure of the adaptive potential of an organism, some kind of 'bacterial IQ'. The slope of the best-fit line on Figure 2a is 2.03, meaning that the total number of signal transduction proteins grows approximately as a square of the genome size. The organisms whose genomes deviate most from this trend can be considered particularly 'smart' or 'dumb' compared to their relatives. There could be different ways to evaluate the relative abundance of signal transduction proteins at the given genome size; the data in Table 1 were calculated using the following formula:

IQ = 5 × 10⁴ (n-5)^1/2 L^-1,

where n is the total number of signal transduction proteins, L is the complete genome size in kb (even counting plasmids, it is a more consistent measure than the number of predicted proteins), 5 × 10⁴ and 5 are arbitrarily chosen empirical coefficients, so that IQ = 100 corresponds to 9 signal transducers in a 1000 kb genome and to 105 transducers in a 5000 kb genome. Accordingly, the IQ value is not defined for organisms with less than 6 signal transduction proteins.

With one exception, all organisms listed in Table 1 are environmental gram-negative bacteria (most gram-positive bacteria and archaea scored much lower) that are highly motile and are known to use a wide variety of electron donors and electron acceptors 73747576. Such versatile organisms as Chromobacterium violaceum, Desulfovibrio vulgaris, Geobacter sulfurreducens, Vibrio vulnificus, and Wolinella succinogenes are also repeatedly found among the leaders in individual categories (Table 2), both in terms of absolute number of signal transduction proteins and of their fraction among all encoded proteins. Remarkably, most of the winners come from the relatively poorly characterized beta-, delta- and epsilon- subdivisions of Proteobacteria. This illustrates the limitations of relying just on Escherichia coli and Bacillus subtilis as model organisms for studying signaling transduction in environmental organisms. The recent efforts on the post-genomic analysis of the versatile gamma-proteobacterium Shewanella oneidensis 77, which encodes a decent set of 46 HKs, 26 MCPs, 7 STYKs, 3 ACs and 52 GGDEF, 28 EAL, and 9 HD-GYP domains [see Additional file 1] might be a step in the right direction. In contrast, E. coli appears to have a relatively low IQ. Although its 30 HKs, 19 GGDEF and 17 EAL domains at first seemed like a high number 1626, E. coli, as well as Salmonella spp. and Yersinia spp., other members of Enterobacteriaceae, looks pretty 'dumb' compared to the representatives of Pseudomonadaceae, Vibrionaceae, or Xanthomonadaceae, particularly with respect to chemotaxis: any sequenced member of the three latter families encodes many more MCPs than the meager 5 MCPs in E. coli. The deep-sea bacterium Idiomarina loihiensis, which belongs to yet another gamma-proteobacterial lineage and whose protein set is just 62% of that of E. coli 78, encodes more diguanylate cyclases and 3 times more MCPs than E. coli. The delta-proteobacterium Bdellovibrio bacteriovorus, a predator that infects an E. coli cell and grows in its periplasmic space, also turned out to have a higher IQ: it has a smaller genome than E. coli but encodes almost twice as many HKs and four times more MCPs.

Histidine kinases are by far the predominant type of sensory proteins (Figure 1), whose distribution in all sequenced organisms generally follows the power law (Figures 3a and 4a). The relative abundance of other types of receptors, however, varies widely among organisms of different phylogenetic lineages (Figure 3b–f). Still, the distribution of their total number also follows power law (Figure 4b). These observations will be analyzed in detail elsewhere. The following is just a brief listing of several unexpected trends:

1. Archaea do not encode AC1- or AC3-type adenylate cyclases, diguanylate cyclases or c-di-GMP-specific phosphodiesterases (with the exception of several highly diverged and probably inactive ORFs), but encode a fair amount of STYKs. In 11 of 20 archaeal genomes, STYKs and class 2 ACs are the only recognizable proteins involved in signal transduction. More than a half of all sequenced archaeal genomes do not encode any MCPs, others encode from 2 to 5 and only the two halophilic species have a large number of MCPs (17 each, Figure 3b).

2. Actinobacteria do not encode MCPs or, for that matter, any other chemotaxis or flagellar proteins (the only one that does, Symbiobacterium thermophilum, probably does not belong to the actinobacterial lineage 79). Instead, actinobacteria encode relatively large numbers of HKs and STYKs (Figure 3a,c). As noted previously, Mycobacterium tuberculosis encodes a relatively high number of AC3s 80, as do two other mycobacteria, M. bovis and M. avium, but not M. leprae (16, 16, 12, and 4, respectively). The regulators of these AC3s remain unknown, although some ACs have been implicated in sensing of the bicarbonate level 81. The dramatically lower number of signaling proteins in M. leprae, compared to other mycobacteria, is in line with the general picture of genome decay in this organism 82.

3. Cyanobacteria encode large numbers of HKs and STYKs, but very few MCPs (e.g. 134, 52 and 3, respectively, in Nostoc PCC 7120 [see Additional file 1]). These data are consistent with previous observations that cyanobacteria encode just several highly conserved MCPs 37 and regulate their motility using HKs (phytochromes) 8384 and ACs 2728.

4. There is great variation between different subdivisions of Proteobacteria with very few common trends. Proteobacteria generally encode few, if any, STYKs, but a large number of MCPs and diguanylate cyclases. The number of ACs is relatively low, except for representatives of the alpha-subdivision. While gamma-proteobacteria typically encode a single AC1 and no more than one AC3, in Pseudomonas aeruginosa this sole AC3 is important for virulence 85.

5. Several bacterial phyla that currently have only a handful of sequenced representatives show highly biased patterns of signal transducer distribution. For example, four sequenced members of the Bacteroidetes (formerly the CFB group) encode a relatively large number of HKs (85 in Bacteroides thetaiotaomicron), but few or no STYKs and no MCPs, ACs or diguanylate cyclases. It would be interesting to see if this trend holds when more genomes of this lineage become available.

The recent genomic data revealed substantial differences in gene content among different strains that, judging by the level of 16S rRNA identity, belong to the same bacterial species 8687. It is therefore not surprising to see dramatic differences in signaling protein content among different species of the same genus. Still, different members of the Bacillus genus show very similar distributions of signaling proteins [see Additional file 1]. In contrast, three sequenced genomes of Clostridium spp. encode dramatically different numbers of MCPs (38 in C. acetobutylicum, 20 in C. tetani and 0 in C. perfringens) and HD-GYP domains (9, 1, and 1, respectively), whereas the content of other signaling proteins is more or less in line with the genome sizes. Accordingly, C. acetobutylicum makes it into the winners list in both MCP and HD-GYP categories (Table 2).

Although not seen in the current data set, domains that are missing in one strain were sometimes found in a different strain of the same species. Thus, although this domain census shows the absence of HD-GYP domains in Yersinia pestis strain CO92 and in Bacillus cereus strain ATCC 14579 [see Additional file 1], this domain is encoded in Y. pestis strain KIM and B. cereus strain ZK. These differences indicate that signaling proteins can be easily acquired and lost, so all observations on the presence or absence of certain signaling system in a certain organism are only as good as the current genome set.

Analysis of complete microbial genomes revealed complex systems of intracellular monitoring that included PAS- and GAF-containing proteins with a variety of output domains 3. The fraction of membrane-bound proteins among all signal transduction proteins encoded in each given genome was evaluated here using three different methods for predicting transmembrane (TM) segments, followed by manual analysis of the outputs. The census showed that while the great majority of HKs and MCPs were membrane-bound, as much as one-third of all HKs and one-sixth of all MCPs did not contain a single TM segment (Figure 1, Additional file 1). In contrast, only about a half of all adenylate and diguanylate cyclases and c-di-GMP phosphodiesterases were membrane-bound; a majority of STYKs and HD-GYP domains were soluble (Figure 1).

It must be noted that not every membrane-bound signal transduction protein is necessarily a sensor of the environmental parameters. An obvious example among HKs is the turgor sensor KdpD, where TM segments serve solely as anchors 88. Aer, the energy-sensing MCP, presents a similar case 89. Conversely, some cytoplasmic sensors might actually sense extracellular signals, e.g. when the sensing domains are present on separate transmembrane polypeptides, as is the case with CheA, the chemotaxis HK. Furthermore, many cytoplasmic sensors respond to signals that are membrane-permeable, such as light, oxygen, H₂O₂; NH₃, and should not be considered purely external or internal. Keeping in mind all these caveats, the predominance of extracellular or intracellular transducers can be used to distinguish organisms that are concerned primarily with sensing environmental parameters ("extroverts") from those more closely monitoring the intracellular milieu ("introverts").

In obligately parasitic bacteria that encode only a handful of signal transduction proteins, most of these proteins are membrane-bound [see Additional file 1]. However, Figure 5a shows that once the total number of signal transduction proteins goes beyond a dozen, the fraction of them that are membrane-bound stabilizes at about 60%, approximately the same in representatives of all phyla, except for Cyanobacteria and Archaea. In the latter two groups, the fraction of membrane-bound signal transduction proteins is close to 30% and also shows very little variance (Figure 5a). Although, as mentioned above, cyanobacteria encode very few MCPs, this fact alone cannot account for the difference between them and all other bacteria. A comparison of other types of signaling proteins encoded in cyanobacteria and proteobacteria (Figure 5b) shows the prevalence of soluble proteins among cyanobacterial HKs, STYKs and GGDEF domains, compared to the prevalence of TM proteins among the same groups of proteins in proteobacteria. The difference between cyanobacterial and archaeal proteins on one hand and proteins from other lineages is most clearly seen in the comparison of HKs (Figure 5c). Remarkably, actinobacteria and firmicutes turn out to be firm extroverts with relatively few intracellular HKs; some of the latter, however, are known to participate in regulation of sporulation 6. This schism between cyanobacteria and all other bacteria with completely sequenced genomes is likely to be due to the much more complex organization of the cyanobacterial cell, which contains intracellular membranes harboring the photosynthetic reaction centers. Among other autotrophic prokaryotes, prevalence of intracellular proteins is seen in methanogenic archaea and in the green sulfur bacterium Chlorobium tepidum, although not in the phototrophic alpha-proteobacterium Rhodopseudomonas palustris [see Additional file 1]. Since archaeal heterotrophs also show low amounts of TM signaling proteins, there does not seem to be a direct connection between 'introvertness' and autotrophic metabolism.

For the purposes of this study, the set of surveyed signal transduction proteins has been limited to just 7 classes of proteins: histidine kinases, methyl-accepting chemotaxis receptors, Ser/Thr protein kinases, adenylate and diguanylate cyclases, c-di-GMP phosphodiesterases with the EAL domain and predicted phosphodiesterases with the HD-GYP domain [see Additional file 1]. Certainly, this list is far from being complete. In a general sense, any cellular protein that participates in cellular adaptation to the changing environment can be considered part of the signaling machinery. Thus, AraC-type transcription regulator, whose DNA-binding properties are modulated by arabinose binding to its N-terminal domain 52, could also be treated as an intracellular signal transducer. According to a recent study by Ulrich, Koonin, and Zhulin, such 'one-component' signalers comprise a majority of signal transduction systems and were the first to arise in evolution 90. More sophisticated mechanisms of signal transduction include two-component (HK and RR) signal transduction systems and a variety of other signaling systems that have been described only in the past several years (see 23212239 for reviews).

This census considered only dedicated signaling systems that consist of more than two individual components. Therefore, transcriptional regulators, even those of complex domain architecture, were left out (for a comprehensive survey of helix-turn-helix-type (HTH) transcriptional regulators, see 91). I have also left out response regulators, which are typically considered together with HKs. One of the reasons for that was the frequent confusion between three classes of response regulators: (i) the single-domain chemotaxis response regulator CheY that transmits the signal through protein-protein interactions; (ii) the DNA-binding response regulators of the CheY-HTH domain architecture, and (iii) the response regulators with CheY-AC, CheY-GGDEF or CheY-GGDEF-EAL domain architectures, which produce secondary messengers, cAMP and c-di-GMP. Here, various proteins containing AC, GGDEF, EAL or HD-GYP domains have been lumped together, just as the chemotaxis signal transduction kinase CheA is typically treated as sensor kinase, despite being just a transmitter in the signaling cascade going from MCPs to the flagellar motor. This approach differed from that of Ulrich et al. 90, who included diguanylate cyclases and c-di-GMP phosphodiesterases (GGDEF and EAL domains, respectively) into the 'one-component' set.

Another important omission in this survey are Ser/Thr protein phosphatases, which can dephosphorylate STYKs, modulating their activity, and should also be able to dephosphorylate the cellular targets of STYKs. However, several surveys of these enzymes have been published recently 193992, and more are apparently on the way. Due to the difficulties in separating true protein phosphatases from phosphatases of other specificities that often produce false-positive hits I have chosen to exclude them from this survey. Several other systems of the bacterial signal transduction machinery have also been left out. These include (i) Ser/Thr kinases of the bacterial (GHKL) type that regulate the activity of the RNA polymerase sigma subunit; (ii) HPr^Ser kinase/phosphorylase and other components of the bacterial PEP-dependent phosphotransferase systems, which regulate chemotaxis, membrane transport (inducer exclusion), and catabolite repression; (iii) the systems that regulate RNA and protein degradation; and many others. A census of each of these systems could be an interesting project in its own right.

The limited scope of this survey, which did not include the sophisticated sporulation machinery of the firmicutes and certain unique (potentially signaling) archaeal domains, could be a reason why representatives of these two groups have generally scored low in the IQ category. Including those proteins into a future version of this census might partly correct that bias, although that would increase the degree of 'introvertness' among archaea even further.

Even within the limited scope of this survey, there is a lot of space for controversy. There are no clear criteria to decide which proteins should be considered HKs or STYKs and which should be not. Thus, the discrepancies of the results presented here and in the papers by the Mizuno group 51634 can all be attributed to their more conservative approach to defining HKs. The survey by Kim and Forst 17 shows a similar undercount of non-canonical HKs. In contrast, counting STYKs in the KinG database 64 used more permissive criteria than those employed here, which resulted in KDO kinases and other related kinases being counted as STYKs. For other signaling domains, there was much less room for disagreement. The counts of MCPs and ACs, presented here, are very similar to those reported, respectively, by Zhulin 40 and Shenoy and Visweswariah 41. All our data with supporting information are available on a public web site 51, which should provide an easy way to analyze any discrepancies and, if necessary, correct the final count.

It is well known that growth in bacterial genome size is accompanied by accumulation of paralogous protein families, which can be easily seen in lineage-specific expansions of transcriptional regulators, metabolic enzymes, and/or surface proteins 939495. It can be argued therefore that the sheer number of signal transduction proteins encoded in a bacterial genome is hardly a good measure of its IQ, as many of these proteins are closely related paralogs. It would seem, however, that lineage-specific expansions that have been fixed in evolution must be of some value to the host organism. Among metabolic enzymes, there are indications of functional diversification even among close paralogs 96. As for signaling proteins, Valley Stewart and colleagues have shown that NarQ and NarX, two paralogous HKs in E. coli, have similar but non-identical functions in modulating cellular response to nitrate and nitrite 9798. Likewise, out of 12 GGDEF domain-containing proteins – potential diguanylate cyclases – encoded in Salmonella Typhimurium genome, one, AdrA, was found to be primarily responsible for regulating biofilm formation in a complex medium, whereas another, STM1987, was critical for biofilm formation in the nutrient-poor medium 99100. These data show that we should be very careful in assigning the same function even to closely related paralogs. Differential regulation of expression and activity of paralogous signal transduction proteins could be yet another sophisticated mechanism allowing the bacterial cell to fine-tune its response to environmental changes. Therefore, until there is clear evidence that functions of paralogous signal transduction proteins are truly identical, the total number of such proteins remains the best measure of the bacterial IQ.

One of the most significant insights to emerge from comparative genome analysis was the recognition of the vast system of intracellular signaling in bacteria. It became clear that many bacteria encode complex systems of intracellular monitoring whose domain organization is very similar to that used in transmembrane signaling: a sensor domain (typically, PAS and/or GAF), followed by HK, AC, GGDEF or EAL output domains 3. In certain cases, soluble HKs, MCPs, and ACs have been experimentally characterized and shown to be involved in monitoring levels of intracellular ATP, oxygen, CO, bicarbonate, nitrate, reactive nitrogen species, and other metabolites and modulating the cellular response to the changes in these parameters 101102103104105. Some intracellular sensors appeared to be specifically geared towards unusual substrates used by the particular bacterium, such as methanol and formaldehyde in Paracoccus denitrificans and Methylobacterium organophilum 106107. In the recently sequenced genome of Dehalococcoides ethenogenes, a major detoxifier of chlorinated organic pollutants, many soluble HKs were found encoded in close proximity to the genes for reductive halogenases, the enzymes that catalyze the dechlorination reactions 108. It was proposed that these HKs respond to intracellular rather than extracellular stimuli, stimulating the expression of reductive halogenases in response to the presence of their chlorinated substrates 108.

This census shows that intracellular signal transduction proteins comprise a significant fraction of all signal transducers encoded in almost any bacterial genome. However, most of them are still uncharacterized and have yet to be recognized as legitimate members of the bacterial signaling network. The finding that these proteins are abundant in many pathogenic as well as free-living bacteria should help focus the attention of the research community on these novel components of the signal transduction network.

The predominance of intracellular signal transduction proteins in cyanobacteria is in stark contrast with the far smaller proportion of such proteins in other bacterial lineages. There could be several possible reasons for this 'introvertness', all linked to the ability of cyanobacteria to conduct oxygenic photosynthesis. Firstly, cyanobacteria harbor a complex system of intracellular membranes carrying the photosynthetic reaction centers. Intracellular signaling proteins could be needed to control formation and functioning of the photosynthetic system, as well as the transition from phototrophic to heterotrophic metabolism and back. The compartmentalization of the cellular interior probably requires a sophisticated system of monitoring conditions within the individual compartments. Last but not the least, cyanobacteria are unique among (known) prokaryotes in that their cells generate oxygen, which other bacteria try to keep outside the cell. The presence of oxygen affects the redox balance in the cytoplasm and leads to oxidative damage of numerous cellular compounds, including ATP, methionine, cysteine, and many others. It is very likely that numerous intracellular HKs that contain PAS domains are involved in maintaining the constant level of the redox potential in the cyanobacterial cell. Surprisingly, Rhodopseudomonas palustris, an alpha-proteobacterium that is also capable of transition between autotrophic and heterotrophic metabolism, does not appear to be an 'introvert' [see Additional file 1]. Hence, it seems that the trend of autotrophic bacteria and archaea being more of 'introverts' and heterotrophic bacteria being more of 'extroverts' might be biased by the current selection of the completely sequenced genomes. It would be interesting to see whether this trend holds when more genomes of bacterial photo- and chemolitotrophs become available.

The knowledge of the phylogenetic distribution of signal transduction systems allows a better understanding of their evolution. Previous analysis of HKs and RRs by Koretke and colleagues led to the conclusion that two-component systems originated in bacteria and radiated into two other domains of life through multiple events of horizontal gene transfer 32. HKs and STYKs appear to be the principal signal transduction proteins in archaea, suggesting that these two classes of proteins could be already present in the last common ancestor of all living organisms (LUCA, 92109110). The absence of AC3-type adenylate cyclases, diguanylate cyclases and c-di-GMP phosphodiesterases in any of the sequenced archaeal genomes is quite remarkable. In fact, the only full-size archaeal AC3 domain known to date has been found in an uncultivated psychrophilic crenarchaeote that exhibited numerous cases of horizontal gene transfer 111. Most archaea, however, encode ACs of class 2 (COG1437 [http://www.ncbi.nlm.nih.gov/COG/new/release/cow.cgi?view=1&cog=COG1437]), which are found in only a handful of organisms outside Archaea 30. These data show that although cAMP is a truly universal second messenger, different domains of life utilize different enzymes for its production and probably employ entirely different mechanisms of cAMP-dependent signaling.

Another remarkable example is the diversity of outputs of the chemotaxis machinery. Although all MCPs counted in this work are very similar, it has been noted 112 that chemotactic signals in diverse bacteria and archaea are being transduced to at least three different motility apparata: the bacterial flagellum, the archaeal flagellum that is unrelated to the bacterial one 113, and to the type IV pili, which are responsible for gliding motility of cyanobacteria and certain other bacteria 84114.

In general, variability of signal transduction protein content in closely related bacteria, uneven distribution of these proteins among well-established phylogenetic lineages, and the presence in many genomes of tight clusters of closely related paralogs indicate that signaling proteins can be easily acquired and lost. Lineage-specific gene duplication and gene loss and lateral gene transfer probably play a key role in shaping the signaling protein repertoire of each given organism. Why, then, would the total number of signal transduction proteins grow as a square of the genome size (Figure 2a) across a wide variety of microorganisms with diverse lifestyles, phylogenetic affinities, and metabolic capabilities? It is tempting to suggest that there must be an underlying mechanism supporting this correlation. For example, the power-law distribution of HKs (Figure 4a) might stem from the simple fact that the number of binary interactions grows as a square of the number of interacting components 115, so that the number of sensory proteins that manage the linearly growing number of metabolic enzymes has to grow as a square of that number. This explanation is somewhat similar to the one offered by van Nimwegen to explain his observation that the number of transcriptional regulators in bacteria also grows as a square of the genome size 71, although his analysis did not include two-component systems. This was also the rationale behind the decision to measure bacterial IQ as a square root, rather than a linear function, of the total number of encoded signal transduction proteins (see the Results section). However, HKs comprise only but a half of all signal transduction proteins counted in this work [see Additional file 1]. The distribution of other types of signal transducers is even more fascinating: while distribution of each individual class of proteins seems almost random (Figures 3b–f), their total number still grows approximately as a square of genome size (Figure 4b). One could speculate that this quadratic dependence determines a near-optimal number of signal transducers at a given genome size. This would mean that during their adaptation to different ecological niches, bacteria evolve to rely primarily on certain types of signal transduction, while other types of transducers can be lost (or not fixed in the genome when acquired by lateral gene transfer). For example, during the reductive evolution of chlamydia, HKs and STYKs were retained, while all other transducers and were lost [see Additional file 1]. In contrast, spirochetes held on to their chemotaxis transducers but mostly lost their STYKs. The recent evidence for non-canonical roles of signal transduction proteins, e.g. regulation of gene expression by the chemotaxis system 116 and regulation of chemotaxis by adenylate cyclases 28, suggests that there is certain flexibility in functions of different transducers that could be used by bacterial evolution to generate even greater diversity of signal transduction mechanisms.

The goal of genome analysis is to predict the organism's physiology and behavior based solely on the genomic sequence. There has been great progress in predicting metabolic pathways 110117118; deciphering signaling pathways so far has lagged behind. Accumulation of complete genome sequences has led to the delineation of many new signaling and signal transduction domains and caused a revolution on our understanding of bacterial regulatory networks 2320119.

I believe that, despite all its limitations, this census would be useful for microbiologists, at least by highlighting still unresolved problems in prokaryotic signal transduction. This work should be complemented by surveys of other components of the signal transduction machinery, including various response regulators, Ser/Thr protein phosphatases, PTS proteins, and many others. Genomes of several environmental microorganisms, including 9-Mb genomes of Myxococcus xanthus, Rhodococcus sp., and Gemmata obscuriglobus, have been completed and are expected to be publicly released in the near future. Owing to their sheer size, these genomes are likely to bring new signaling domains and illuminate even more regulatory relations. Myxococcus xanthus, which reportedly encodes close to 200 HKs and many STYKs, would probably become a leader in both these categories.

The example of M. xanthus exposes certain flaws in the IQ calculation method used in this work. This bacterium has extremely complex behavioral patterns 114, but, at 9.1 Mb, it would need to encode more than 550 signal transduction proteins just to make it into the winners' list (Table 1). Certainly, better ways to evaluate bacterial IQ are needed, but that should be subject of a future work. Still, I believe that in the era of 'systems biology' when cellular metabolic pathways are being routinely modeled on a whole-genome level 50120 and the cell itself is treated more as a machine with a number of interacting parts 121122, it is important to keep in mind the real complexity of the signal network encoded in each given prokaryotic genome and have an easy measure of this complexity.

I also hope that this census will help us get a better understanding of the microbial diversity and the unique ways that bacteria use to adapt to changing environment. Such understanding is becoming increasingly important as our earlier methods of controlling bacterial growth with one-size-fits-all wide-spectrum antibiotics show progressively diminishing results.

Complete genome sequences of 167 bacterial and archaeal species, sequenced by the end of 2004, were downloaded from the NCBI's Genomes database 123 or searched directly through the NCBI web site. Only one representative genome per species was used, usually the first one to be publicly released, according to the NCBI Genomes database listing. Exceptions were made for Escherichia coli, represented by two strains, K12 [GenBank:U00096] and O157:H7 [GenBank:BA000007], and Salmonella enterica, represented by three serovars, Paratyphi [GenBank:CP000026], Typhi [GenBank:AL513382], and Typhimurium [GenBank:AE006468]. For Prochlorococcus marinus, strain CCMP1375 [GenBank:AE017126] genome was used, the middle-sized one of the three. Among other simultaneously released genomes, Staphylococcus aureus N315 [GenBank:BA000018], Streptococcus thermophilus CNRZ1066 [GenBank:CP000024], and Thermus thermophilus HB27 [GenBank:AE017221] genomes were used.

The complete list of histidine kinases was compiled separately for each particular phylum of bacteria from the results of BLAST searches against selected genomes using the NCBI's Genomic BLAST tool 124, followed by iterative PSI-BLAST searches 59. Typically, the searches used as the query sequence the C-terminal fragment (residues 301–579) of the well-characterized histidine kinase PhoR [UniProt:P23545 http://www.expasy.org/cgi-bin/niceprot.pl?P23545] from Bacillus subtilis, which contains both HisKA and HATPase domains 125, and a position-specific scoring matrix (PSSM) derived from an alignment of well-characterized histidine kinases (both available as Supplementary Material). Additional searches against the NCBI's Reference Sequence (RefSeq) database 126127 were performed through the NCBI BLAST web interface http://www.ncbi.nlm.nih.gov/BLAST/ by limiting the search space to the given phylum (e.g. Actinobacteria [orgn]) and excluding reference sequences of incomplete genomes (srcdb_refseq [prop] NOT srcdb_refseq_model [prop]). The PSI-BLAST searches used strict inclusion threshold expect values of 10^-5–10^-7 (adjusting as necessary) and were iterated until no newly retrieved sequences belonged to HKs. The total numbers of copies of each signaling domain encoded in each given genome were estimated using the "Taxonomy Report" option in the BLAST output. Potential false-positive hits were checked at every step of PSI-BLAST using the CDD Domain viewer and manually removed (unselected) from the hit list for the next iteration of PSI-BLAST. In each case where the HATPase domain was easily recognized but HisKA domain was not, a BLAST2sequences 128 search was performed to check whether the HATPase domain was preceded by a conserved region carrying a conserved His residue. The presence of such His-containing regions would indicate that those questionable proteins (e.g., mlr1749 [UniProt:Q98JW4 http://www.expasy.org/cgi-bin/niceprot.pl?Q98JW4_RHILO] and other members of COG3920 [http://www.ncbi.nlm.nih.gov/COG/new/release/cow.cgi?view=1&cog=COG3920]) comprise legitimate HKs, contrary to the view of Hagiwara et al. 34.

Alternatively, PSI-BLAST searches were run against a local copy of the RefSeq database, using the same query sequence and search parameters with additional filtering against sequences translated from unfinished genomes (ZP_xxxxxxxx entries). The resulting hits were compared against the NCBI Taxonomy database to ensure that they all came from a single organism (only one genome of each bacterial species, usually the first one to be sequenced, was used in this analysis). Similar protocol was used to search for histidine kinases in other bacterial phyla.

Owing to the relatively high sequence conservation of the MCP, ACyc, GGDEF, and EAL domains, manual checking of the PSI-BLAST outputs revealed very few false-positive hits. In the case of the two latter domains, many low-scoring proteins had numerous amino acid changes, including ones in the likely active sites (see 22267). No attempt has been made to sort these domains into active and inactive ones. For the HD-GYP domain, which comprises a typical HD superfamily phosphoesterase domain with a number of additional conserved residues, high-scoring BLAST hits to the standard HD domains were filtered based on the shorter length of those hits.

Transmembrane (TM) segments in verified sets of signal transduction proteins from various phylogenetic lineages were predicted using PHDhtm 129 and TMHMM 130 programs. The results were sorted into three bins: TM proteins (≥ 2 TM segments), 1 TM proteins, and soluble proteins, and the discrepancies between predictions of the two programs were manually inspected. Comparison of the results revealed many false-negative assignments, so that prediction of a TM segment by either program typically turned out to be justified. Questionable cases were also checked using the HMMTop 131 program, which, however, produced both false-negative and false-positive predictions of TM segments. Therefore, HMMTop assignments were considered only when supported by either PHDhtm or TMHMM results.