Genome sequences, predicted ORFs and annotation files of the following 29 cyanobacterial genomes were downloaded from the NCBI website at ftp://ftp.ncbi.nlm.nih.gov/genomes/Bacteria: Acaryochloris marina MBIC11017 (MBIC11017), Anabaena variabilis ATCC 29413 (ATCC29413), Anabaena sp. PCC 7120 (PCC7120), Gloeobacter violaceus PCC 7421 (PCC7421), Prochlorococcus marinus str. AS9601 (AS9601), Prochlorococcus marinus str. MIT 9211 (MIT9211), Prochlorococcus marinus str. MIT 9215 (MIT9215), Prochlorococcus marinus str. MIT 9301 (MIT9301), Prochlorococcus marinus MIT9303 (MIT9303), Prochlorococcus marinus MIT9312 (MIT9312), Prochlorococcus marinus str. MIT 9313 (MIT9313), Prochlorococcus marinus str. MIT 9515 (MIT9515), Prochlorococcus marinus str. NATL1A (NATL1A), Prochlorococcus marinus str. NATL2A (NATL2A), Prochlorococcus marinus CCMP1375 (CCMP1375), Prochlorococcus marinus MED4 (MED4), Synechococcus elongatus PCC 6301 (PCC6301), Synechococcus elongatus PCC 7942 (PCC7942), Synechococcus sp. CC9311 (CC9311), Synechococcus sp. CC9605 (CC9605), Synechococcus sp. JA-2-3B'a(2–13) (B-Prime), Synechococcus sp. JA-3-3Ab (A-Prime), Synechocystis sp. PCC 6803 (PCC6803), Synechococcus sp. CC9902 (CC9902), Synechococcus sp. RCC307 (RCC307), Synechococcus sp. WH 7803 (WH7803), Synechococcus sp. WH8102 (WH8102), Thermosynechococcus elongatus BP-1 (BP-1), and Trichodesmium erythraeum IMS101 (IMS101).

We predicted operon structures in each cyanobacterial genomes using the Operon Finder Software (OFS) developed by Westover et al 34. OFS predicts operons based on three pieces of information, including the intergenic distance, functional relatedness of gene annotations, and conserved gene neighborhoods. In this study, both a multi-gene operon and a singleton operon containing only one gene are referred as a transcription unit (TU).

A simple bidirectional best hit (BDBH) approach using the BLASTP program with an E-value cutoff 10^-10 was used for the prediction of orthologous genes between each pair of genomes.

To construct the CRP tree, the full length amino acid sequences of cyanobacterial CRPs were identified by the criteria described above using SyCRP1 (sll1371) in PCC6803 as the query sequence. Multiple sequence alignments of the identified cyanobacterial CRP sequences and that of E. coli K12 were performed using ClustalW implemented in MEGA 35 with default settings. A neighbor-joining (NJ) tree with Poisson correction was constructed using the MEGA program with the E. coli CRP (GI:16131236) being the outgroup. To construct the tree in Figure S1 (see Additional file 1), we used SyCRP1 as the query sequence to search the RefSeq database using BLASTP with an E-value cutoff 10^-7. If there were multiple hits from a species, the hit with the smallest E-value was identified as the CRP in that species. The resulting sequences were used to construct an un-rooted tree in the similar way as described above. To construct the species tree, the DNA sequences of the 16S rRNA genes of the sequenced cyanobacteria and that of E. coli K12 were aligned using ClustalW with manual refinement. After the indels were discarded, the final alignments contain 1311 positions. A Neighbor-Joining tree of the 16 rRNA gene sequences was constructed with the E. coli K12 sequence being the outgroup using Kimura 2-parameter model. Statistical significance at each node in the trees was evaluated using 1000 bootstrap resamplings.

A previous screening of the target genes of SyCRP1 in PCC6803 using microarray showed that 18 genes in 13 putative TUs were down-regulated in a sycrp1 disruptant compared with the wild type strain 25. Our preliminary manual scanning of the upstream promoter regions of these TUs revealed that four of them contain a putative CRP binding site similar to those in E. coli (the first four TUs in Table 1). The rest nine TUs that do not contain putative CRP binding sites are likely to be regulated indirectly by CRP through different regulators. In addition, since the crp gene is auto-regulated in E. coli as well as in many other species 36, we also manually scanned the upstream inter-TU region of sycrp1 (sll1371) and identified a putative CRP binding site in it (Table 1). We thus used the orthologues (if they exist and were identified by the BDBH method) of these five TUs for the phylogenetic footprinting of CRP binding sites in cyanobacterial genomes.

For this purpose, we pooled the entire upstream inter-TU regions of these five TUs in PCC6803, as well as those of TUs containing their orthologous genes in other cyanobacterial genomes which encode at least one crp gene. Motif finding programs including CUBIC 37 and BioProspector 38, were then applied to predict CRP binding sites in these sequences. CUBIC is a graph theoretic based algorithm that identifies highly similar k-mers in a set of pooled sequences; while BioProspector uses a Gibbs sampling strategy to find overrepresented k-mers in a set of pooled sequences. Putative CRP binding sites with high scores were manually picked up from the motif finding results, and were used to build a preliminary profile of the CRP binding sites in cyanobacterial genomes.

Since the number of binding sites used to construct the preliminary profile of CRP binding sites was relatively small (for the reason, see the Results section), in order to minimize possible bias of binding site sampling, we conducted a one-round iteration to obtain a more representative profile of the CRP binding sites. To this end, we scanned cyanobacterial genomes with the preliminary profile using the techniques described below, and picked the high scoring sites from each genome to construct the more representative profile of the CRP binding sites. We then used this final profile for genome-scale predictions of CRP binding sites in the cyanobacterial genomes.

The whole genome screening of all possible CRP binding sites were performed using an algorithm that we have developed previously 32. The design of this algorithm is to enhance the prediction specificity by integrating the information of co-occurrence of multiple binding sites in the upstream region of a gene and that in the upstream regions of its orthologues in related genomes. Briefly, the final profile of the CRP binding sites obtained above was first used to scan all the inter-TU regions of the cyanobacterial genomes. The best motif was returned for each inter-TU region. Then the 19 to 31 bp downstream region of each putative binding site was further scanned for an E. coli -10 σ^-70 like box (TAN₃T) using a corresponding profile that we have constructed previously from cyanobacterial genomes 32, to which a σ-factor of RNA polymerase is likely to bind to transcribe the downstream TU. Then, the upstream inter-TU regions of the orthologues of the genes in that particular TU in other cyanobacterial genomes were scanned for similar CRP binding sites and -10 like boxes. A score that combines these three pieces of information was computed to rank the putative CRP binding sites for each possible CRP-regulated TU.

Specifically, let M be the final profile of the CRP binding sites obtained above. For each predicted transcription unit U(g₁, ..., g_n) containing genes g₁, ..., g_n in a genome G, we extract the entire upstream intergenic region (if that region is larger than 800 bp, then only the 800 bp upstream the translation starting site were extracted) and denoted it as IU(g1,...,gn) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemysaK0aaSbaaSqaaiabdwfavjabcIcaOiabdEgaNnaaBaaameaacqaIXaqmaeqaaSGaeiilaWIaeiOla4IaeiOla4IaeiOla4IaeiilaWIaem4zaC2aaSbaaWqaaiabd6gaUbqabaWccqGGPaqkaeqaaaaa@39E0@. We also extracted a random DNA sequence with the same length as IU(g1,...,gn) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemysaK0aaSbaaSqaaiabdwfavjabcIcaOiabdEgaNnaaBaaameaacqaIXaqmaeqaaSGaeiilaWIaeiOla4IaeiOla4IaeiOla4IaeiilaWIaem4zaC2aaSbaaWqaaiabd6gaUbqabaWccqGGPaqkaeqaaaaa@39E0@ from the coding region, denoted as CU(g1,...,gn) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4qam0aaSbaaSqaaiabdwfavjabcIcaOiabdEgaNnaaBaaameaacqaIXaqmaeqaaSGaeiilaWIaeiOla4IaeiOla4IaeiOla4IaeiilaWIaem4zaC2aaSbaaWqaaiabd6gaUbqabaWccqGGPaqkaeqaaaaa@39D4@. We say that IU(g1,...,gn) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemysaK0aaSbaaSqaaiabdwfavjabcIcaOiabdEgaNnaaBaaameaacqaIXaqmaeqaaSGaeiilaWIaeiOla4IaeiOla4IaeiOla4IaeiilaWIaem4zaC2aaSbaaWqaaiabd6gaUbqabaWccqGGPaqkaeqaaaaa@39E0@ and CU(g1,...,gn) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4qam0aaSbaaSqaaiabdwfavjabcIcaOiabdEgaNnaaBaaameaacqaIXaqmaeqaaSGaeiilaWIaeiOla4IaeiOla4IaeiOla4IaeiilaWIaem4zaC2aaSbaaWqaaiabd6gaUbqabaWccqGGPaqkaeqaaaaa@39D4@ are associated with U(g₁, ..., g_n) and with each genes g₁, ..., g_n as well. All the extracted IU(g1,...,gn) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemysaK0aaSbaaSqaaiabdwfavjabcIcaOiabdEgaNnaaBaaameaacqaIXaqmaeqaaSGaeiilaWIaeiOla4IaeiOla4IaeiOla4IaeiilaWIaem4zaC2aaSbaaWqaaiabd6gaUbqabaWccqGGPaqkaeqaaaaa@39E0@ from one genome are denoted as set I_U, and all the CU(g1,...,gn) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4qam0aaSbaaSqaaiabdwfavjabcIcaOiabdEgaNnaaBaaameaacqaIXaqmaeqaaSGaeiilaWIaeiOla4IaeiOla4IaeiOla4IaeiilaWIaem4zaC2aaSbaaWqaaiabd6gaUbqabaWccqGGPaqkaeqaaaaa@39D4@ in the same genome are denoted as set C_U. For each t ∈ I_U or t ∈ C_U, we scan for possible CRP bindings sites using the profile M. The score of a putative binding site found in t by scanning with profile M is defined as,

s M ( t ) = max ⁡ h ⊂ t ∑ i = 1 l I i ln ⁡ p ( i , h ( i ) ) q ( h ( i ) ) , MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4Cam3aaSbaaSqaaiabd2eanbqabaGccqGGOaakcqWG0baDcqGGPaqkcqGH9aqpdaWfqaqaaiGbc2gaTjabcggaHjabcIha4bWcbaGaemiAaGMaeyOGIWSaemiDaqhabeaakmaaqahabaGaemysaK0aaSbaaSqaaiabdMgaPbqabaaabaGaemyAaKMaeyypa0JaeGymaedabaGaemiBaWganiabggHiLdGccyGGSbaBcqGGUbGBjuaGdaWcaaqaaiabdchaWjabcIcaOiabdMgaPjabcYcaSiabdIgaOjabcIcaOiabdMgaPjabcMcaPiabcMcaPaqaaiabdghaXjabcIcaOiabdIgaOjabcIcaOiabdMgaPjabcMcaPiabcMcaPaaakiabcYcaSaaa@5B65@

I i = ( ∑ b ∈ { A , C , G , T } p ( i , b ) ln ⁡ p ( i , b ) q ( b ) ) / a , MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemysaK0aaSbaaSqaaiabdMgaPbqabaGccqGH9aqpcqGGOaakdaaeqbqaaiabdchaWjabcIcaOiabdMgaPjabcYcaSiabdkgaIjabcMcaPiGbcYgaSjabc6gaULqbaoaalaaabaGaemiCaaNaeiikaGIaemyAaKMaeiilaWIaemOyaiMaeiykaKcabaGaemyCaeNaeiikaGIaemOyaiMaeiykaKcaaaWcbaGaemOyaiMaeyicI4Saei4EaSNaemyqaeKaeiilaWIaem4qamKaeiilaWIaem4raCKaeiilaWIaemivaqLaeiyFa0habeqdcqGHris5aOGaeiykaKIaei4la8IaemyyaeMaeiilaWcaaa@58AC@

a = n + 1 n + 4 ln ⁡ ( n + 1 ) − ln ⁡ ( n + 4 ) − 1 n + 4 ∑ b ∈ { A , C , G , T } ln ⁡ q ( b ) − n n + 4 ln ⁡ min ⁡ b ∈ { A , C , G , T } q ( b ) , MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@8148@

where l is the length of the binding sites of M, h any substring of t with length l, h(i) the base at the i-th position of h, p(i, b) the relative frequency of base b occurring at position i in M, q(b) the background frequency of base b, and n the number of sequences used to construct M. When computing p(i, b), a pseudo count 1 is added to the frequency of each base at each position, and a is for normalization to keep I_i within the range of [0,1].

When multiple profiles are considered for scanning t, we sum up the individual S_M(t) s as defined by

s M 1 ... M z ( t ) = ∑ j = 1 z s M j ( t ) . MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4Cam3aaSbaaSqaaiabd2eannaaBaaameaacqaIXaqmaeqaaSGaeiOla4IaeiOla4IaeiOla4Iaemyta00aaSbaaWqaaiabdQha6bqabaaaleqaaOGaeiikaGIaemiDaqNaeiykaKIaeyypa0ZaaabCaeaacqWGZbWCdaWgaaWcbaGaemyta00aaSbaaWqaaiabdQgaQbqabaaaleqaaOGaeiikaGIaemiDaqNaeiykaKcaleaacqWGQbGAcqGH9aqpcqaIXaqmaeaacqWG6bGEa0GaeyyeIuoakiabc6caUaaa@4940@

For this study, we use M₁ for the CRP binding sites and M₂ for the -10 like box (TAN₃T).

For the inter-TU sequence t associated with U(g₁, ..., g_n) in genome G, if g_i has orthologues in m_i closely related genomes G1,...,Gmi MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4raC0aaSbaaSqaaiabigdaXaqabaGccqGGSaalcqGGUaGlcqGGUaGlcqGGUaGlcqGGSaalcqWGhbWrdaWgaaWcbaGaemyBa02aaSbaaWqaaiabdMgaPbqabaaaleqaaaaa@36B7@, we denote o_k(g_i) as the inter-TU sequence upstream the TU containing g_i 's orthologue in genome G_k. When the presence of similar motifs in o_k(g_i) is also considered, the score of co-occurrence of the multiple binding sites in t is redefined as

s ( t ) = s M 1 ... M z ( t ) + max ⁡ 1 < i < n ∑ j = 1 z ∑ k = 1 m i l j − d i , j , k m i l j s M j ( o k ( g i ) ) , MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@78FD@

where d_{i, j, k} is the Hamming distance between the sequence found by using profile M_j in t and o_k(g_i), l_j the length of binding site motif with profile M_j.

We evaluated the statistical significance of our predictions by comparing the probability of finding a high scoring binding site in inter-TU sequences with that of finding the same high scoring binding site in randomly extracted coding sequences 3233. Let p(S_t>s) be the probability that an extracted sequence t (t ∈ I_U or t ∈ C_U) contains a putative binding site with a score (S_t) larger than s. To avoid possible biased sampling, 300 C_us for each TU in each genome were randomly generated, and S_Cu was computed for each C_u. Then a log-odd ratio (LOR) function defined as following was used to estimate the confidence of the prediction:

L O R ( s ) = ln ⁡ p ( S I U > s ) p ( S C U > s ) . MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemitaWKaem4ta8KaemOuaiLaeiikaGIaem4CamNaeiykaKIaeyypa0JagiiBaWMaeiOBa4wcfa4aaSaaaeaacqWGWbaCcqGGOaakcqWGtbWudaWgaaqaaiabdMeajnaaBaaabaGaemyvaufabeaaaeqaaiabg6da+iabdohaZjabcMcaPaqaaiabdchaWjabcIcaOiabdofatnaaBaaabaGaem4qam0aaSbaaeaacqWGvbqvaeqaaaqabaGaeyOpa4Jaem4CamNaeiykaKcaaiabc6caUaaa@4A9D@

Since p(SCU>s) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemiCaaNaeiikaGIaem4uam1aaSbaaSqaaiabdoeadnaaBaaameaacqWGvbqvaeqaaaWcbeaakiabg6da+iabdohaZjabcMcaPaaa@3546@ is the probability of type-I error for testing the null hypothesis that I_u does not contain a binding site when SIU MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4uam1aaSbaaSqaaiabdMeajnaaBaaameaacqWGvbqvaeqaaaWcbeaaaaa@2FB6@ is greater than a cutoff s, we used it to estimate the false positive rate of the prediction results. In this way, p(SCU>s) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemiCaaNaeiikaGIaem4uam1aaSbaaSqaaiabdoeadnaaBaaameaacqWGvbqvaeqaaaWcbeaakiabg6da+iabdohaZjabcMcaPaaa@3546@ could be also considered as an empirical p-value, and a cutoff of p < 0.01 was used for the CRP binding site prediction in each genome.

Using the BDBH algorithm and the criterion described in Methods, we identified orthologues of SyCRP1 (sll1371) of PCC6803 in 12 of the 29 sequenced cyanobacterial genomes. All these 12 genomes encode one SyCRP1 homologue, with the exception that the PCC6803 and PCC7120 genomes contain two: sll1371 and sll1924 (SyCRP2) in PCC6803, and alr0295 (AnCRPA) and alr2325 (AnCRPB) in PCC7120, which are identified by a unidirectional BLASTP search using SyCRP1 of PCC6803 as the query and an E-value cutoff 10^-20. The phylogenetic tree of these 12 CRP orthologues shows that they fall into two distinct groups (Figure 1a). There are clear subtle differences between the DBDs of the two groups. Nonetheless, the residues of the DBD that are in direct interaction with the DNA counterpart as revealed by the crystal structure of E. coli CRP/DNA complex 39, are highly conserved in all these cyanobacterial CRP sequences, suggesting that they might bind to similar DNA sequences (Figure 1b).

When the 18 genes located in the five TUs (Table 1) in PCC6803 that are likely to be regulated by SyCRP1 in this species, are searched against the 12 cyanobacterial genomes that encode crp genes, we found a total of 30 orthologues in five genomes. These 30 orthologous genes are located in a total of 10 TUs in these five genomes, indicating that these TUs are not well conserved in these 12 genomes. By using the phylogenetic footprinting techniques (see Methods), we predicted a total of eight putative CRP binding sites from the 10 upstream inter-TU regions, suggesting that the CRP binding sites are largely shared by these orthologous genes. In order to increase the representation of the profile of the CRP binding sites and to minimize the possible bias of our original choice of the five TUs, we performed a one-round iteration of putative CRP binding site scanning using this preliminary profile constructed from these eight putative CRP binding sites (see Methods). From this preliminary whole genome scanning results, we selected a total of 112 putative CRP binding sites (Table S1 in Additional file 2) with high scores to construct the final profile of CRP binding sites. These sites display a strong pseudo-palindromic structure with consensus TGTGAN₆TCACA (Figure 1c), which is similar to the canonical CRP binding sites in E. coli, suggesting that the pattern of CRP binding sites is well conserved between cyanobacterial and E. coli. This result is also in agreement with the observation that the binding sites of members of the CRP/FNR superfamily maintain a high level of conservation across difference lineages 40.

We then apply our motif scanning algorithm to predict putative CRP binding sites in the 12 cyanobacterial genomes that encode a crp gene using the profile of the 112 putative CRP binding sites as well as that of the previously prepared -10 like box from cyanobacteria 32. The log-odds ratio (LOR) function of the predictions in each genome is all high when the score s is high (Figures 2), suggesting that these genomes are likely to contain functional CRP binding sites. The predictions in each of the 12 genomes at p-value < 0.01 are listed in Tables S3-S14 in Additional file 2, and all the prediction results are available upon request. In general, all these predicted CRP promoters contain a pseudo-palindromic CRP binding site and most of them also contain a downstream TA-rich σ-factor binding site, therefore they are likely to be true CRP promoters. As shown in Table S12 (see Additional file 2), the five CRP binding sites associated with the five TUs (shown in bold) in PCC6803, which we selected as the starting point of the our genome-wide prediction, were not necessarily ranked very high (they were ranked 4th, 12th, 14th, 17th, and 43rd) among the 59 predicted CRP binding sites in that genome, suggesting that our initial choice of the five TUs did not bias our prediction to them. Table 2 summarizes the predictions in the 12 cyanobacterial genomes, and the most prevalent genes of predicted CRP regulons are listed in Table S2 in Additional file 2.

Although it has been estimated that CRP controls the expression of more than 200 TUs in E. coli 2, and the RegulonDB (release 5.8) contains 288 experimentally verified CRP binding sites, the number of predicted CRP binding sites at p-value < 0.01 in each cyanobacterial genomes is relatively small, ranging from 29 in BP-1 (containing 1075 putative TUs) to 249 in ATCC29413 (containing 3300 putative TUs), if one considers that the E. coli K12 genome encodes a total of 2070 putative TUs (predicted by the algorithm described in Methods), and that some of these cyanobacterial genomes contain much more TUs/genes (Table 2) than the E. coli K12 genome does. This result suggests that cyanobacterial CRPs might regulate fewer genes than the E. coli CRP does. We then ask whether the target genes of CRPs are conserved between E. coli and cyanobacteria, as well as among the 12 cyanobacterial genomes, given that the DBDs of CRPs as well as their binding sites are highly conserved.

Although these 12 cyanobacterial genomes share 13.88~32.19% of their genes with E. coli (Table 2), they almost have no common genes in their CRP regulons with that of E. coli K12 (Tables 3, and S3-S14 in Additional file 2), suggesting that the CRPs in cyanobacteria have adapted to regulate very different sets of genes than those in E. coli during the course of evolution. On the other hand, the majority of the CRP targets in these cyanobacterial genomes are not conserved either, rather, only a small portion of the CRP targets are shared by more than 2 of the 12 cyanobacterial genomes (Figure 3), even though our motif scanning algorithm might be biased toward the genes that have orthologues in reference genomes (see Methods). In the extreme cases, many CRP targets are species or lineage specific, suggesting that the targets of CRP have changed very rapidly since the speciation of these cyanobacterial genomes. For example, only 32 out of the 149 putative CRP target genes in the PCC6803 genome are conserved in at least 2 of the 12 cyanobacterial genomes. The only exception is the PCC7120 genome in which the number is 218 out of 442. However, this is largely because the reference genomes include a closely related species ATCC29413. Moreover, the most conserved putative CRP-regulated genes are only shared by five genomes (Figure 3). In addition, 17 of the 29 sequenced cyanobacterial genome do not encode the crp gene, suggesting that CRP is not required for the life of these 17 organisms; alternatively, the function of CRP in these organisms have been replaced by another TF.

Based on the functional annotations of the CRP target genes that we have predicted in this study, CRP seems to be involved in a rather diverse spectrum of functions in cyanobacteria (Table 4) as summarized below.

The presence of crp genes in some, but not all cyanobacterial genomes, raised the question about their evolutionary origin: are crp genes in cyanobacteria acquired through horizontal gene transfer (HGT) from other species, or are they vertically inherited from their last common ancestor, but lost in some cyanobacterial species/strains? To address this question, we constructed a species tree of the 29 sequenced cyanobacterial genomes based on their 16S rRNA gene sequences (Figure 4). Although some nodes in the tree are not well supported by the currently available sequence data, the relationships of the species in the tree are in excellent agreement with a previously inferred cyanobacterial species tree 44. As shown in Figure 4, the genomes encoding a crp gene do not form a particular monophyletic group, rather, they are sporadically scattered across the tree. On the other hand, all the known cyanobacterial CRPs form a monophyletic group (Figure S1 in Additional file 1), suggesting that they are likely derived from a common ancestor. Therefore, the current sporadic distribution of the crp genes within the sequenced cyanobacteria likely resulted from differential gene losses during the course of evolution. Independent acquisition of the crp gene by individual cyanobacterial lineage/species, though theoretically possible, is less parsimonious since it would entail multiple HGT events from the same or closely related donors. Therefore, we conclude that the 17 cyanobacterial genomes that do not encode a crp gene actually lost their original ones during the course of evolution to adapt to their respective environments. It is interesting to note that most of the species lacking a crp gene belong to marine cyanobacteria; these species often have a reduced genome and inhabit a relative stable oligotrophic environment. On the other hand, the species that contain a crp gene are distributed in both fresh water and terrestrial environment. This again suggests that the crp gene might be beneficial to the species that live in a variable environment, and those that have adapted to a more stable environment such as oligotrophic ocean lost their crp genes.

We have previously shown that when a genome lost a TF in the course of evolution, then it would rapidly lose its cognate biding sites in inter-TU regions 3233. To extend this conclusion, we applied our CRP binding site prediction algorithm to the 17 cyanobacterial genomes that do not encode a CRP orthologue. Indeed, in four genomes, namely, CC9902, RCC307, WH8102 and WH7803 (Figures S2 in Additional file 3), the LOR function oscillates around zero as the score s increases, indicating that there is no significant difference between the signal of CRP binding sites in the inter-TU regions and that in the randomly selected coding regions (see Methods). These results strongly suggest that these four genomes are unlikely to contain functional CRP binding sites, which is in agreement with our previous observation 3233. However, there is a clear CRP binding site signal in the rest of 13 genomes as indicated by their relatively high LOR when the score s is high, suggesting that there exist CRP-like binding sites in the inter-TU regions in these genomes. The reason for this unexpected observation is unknown, but one explanation would be that these sites are bound by a different regulator in these genomes. To identify possible TFs in these genomes that are likely bind these CRP-like binding sites, we analyze the distribution of the CRP/FNR superfamily in these genomes, and found that they all encode at least one member of the superfamily, which are likely to recognize these CRP-like binding sites, as it has been shown that members of the CRP/FRN superfamily recognize similar consensus sequence, and the binding specificity is achieved through competitive binding among the members.

Studies have shown that CRP in E. coli functions as an important global regulator controlling the expression of genes involved in many pathways such as the carbon and energy metabolism pathways. It was also reported that CRP regulates a variety of genes in other bacterial species. For instance, one recent study showed that CRP-like protein regulates genes involved in quorum sensing, motility and intestinal colonization in Vibrio cholerae 7. In this study, we suggest that CRP in cyanobacteria seems to have distinct functions. First, our results show that the members of the CRP regulons in cyanobacteria have little in common with those in E. coli (Table 3), which is consistent with the observation that genes whose expression is mostly affected by sycrp1 disruption in PCC6803 are involved in the type IV pilus synthesis 82529. In contrast, genes that are involved in carbon and energy metabolisms as seen in E. coli are not significantly affected by sycrp1 disruption 82529. Second, cyanobacterial CRPs seem to regulate distinct sets of genes specific to each lineage or strain as we can not clearly identify a particular set of genes common to most of the 12 cyanobacterial genomes that are under CRP control (Figure 3, Table S2 in Additional file 2). Third, more than half (17) of the 29 completely sequenced cyanobacterial genomes do not encode a CRP orthologue, suggesting that CRP is dispensable in these strains/species. For a closely related group of species, it is unlikely that an essential global regulator can be replaced or lost in some genomes while present in the others. Therefore, we conclude that CRP is not a global regulator with conserved functions; instead, it is likely a lineage-specific global regulator.

In fact, it has been shown that global regulators are not necessarily conserved in moderately related species. For instance, among the 7 and 6 global regulators in E. coli and B. subtilis, respectively, none of them is in common 11. Furthermore, it is not surprising that CRP in cyanobacteria does not function as a conserved global regulator as seen in E. coli, given that CRP is mainly involved in the transcriptional regulation of genes related to carbon metabolism in E. coli, while organic carbon assimilation is no longer a constraint for the growth of autotrophic cyanobacteria. On the other hand, since nitrogen assimilation is a constraint for cyanobacteria, the nitrogen assimilation regulator, NtcA, a member of the CRP/FNR family, which is unanimously encode in all the 29 sequenced genomes, has been characterized as one of the conserved global regulators in cyanobacteria. Thus, it seems that global regulators are often lineage-specific and that the environment plays a vital role in determining which TF functions as a global regulator, and which genes are regulated by the TF.

It has been reported that SyCRP1 regulates the cellular motility in PCC6803, as sycrp1 disruptants were devoid of mobility and showed reduced type IV pilus biogenesis 8. In another study, it was shown that the operons slr1667-slr1668 and slr2015-slr2018 in PCC6803, which are involved in type IV pilus biosynthesis 2945, were down-regulated in sycrp1 disruptants using microarray gene expression profiling 25. In consistent with these findings, we have predicted putative CRP binding sites for these genes. However, the slr1667-slr1668 genes were unique to PCC6803, and the orthologues of slr2015-slr2018 could only be found in MBIC11017 (Table S2 in Additional file 2) among the 29 cyanobacterial genomes. Thus, this function of CRP is likely to be restricted to these two species/strains only. In addition, based on our predictions of CRP regulons in the 12 cyanobacterial genomes, we argue that CRP might be also involved in other functions in different cyanobacterial lineage/strains, including carbon fixation, photosynthesis, nitrogen fixation, ion channels/transporters, and two-component signal transduction, etc. (Table 4, Table S3 – S14 in Additional file 2). Furthermore, as a large portion of our predicted CRP binding sites are associated with hypothetical proteins, CRP might be involved in the regulation of the other novel functions yet to be discovered. In this regard, we have provided a set of candidates for further experimental characterization. However, due to the lack of sufficient information about the functions of the CRP target genes, it is currently rather difficult to derive a general pathway model involving CRP regulons in cyanobacteria. Lastly, AnCrpA was shown to regulate the expression of genes related to nitrogen fixation in PCC7120 28, including all1517 (nifB), all1439, all1432 (hesA), alr2515 (coax-II), and alr2834 (hepC). However, our algorithm failed to find high scoring CRP binding sites for all these genes, the possible reasons for this are addressed below.

Our failure to identify high scoring CRP binding sites in the upstream regions of these nitrogen fixation genes in PCC7120 is actually in agreement with the finding by Suzuki and coworkers who suggested that AnCrpA might have a different binding site pattern from the conventional pseudo-palindromic motif 28. However, an in vitro binding affinity test using EMSA showed that AnCrpA could bind to the conventional palindromic motif 27. Thus, a possible explanation of this inconsistency would be that AnCrpA in PCC7120 could form a monomer or a heterodimer in addition to a homodimer, given that two CRP homologues are encoded in this genome, and one of them does not have a DBD. Such a monomer or heterodimer might favor a non-canonical CRP binding site, while the homodimer remains its ability to bind to the conventional palindromic motif. Clearly, a more in-depth study on this topic is needed to verify this hypothesis. This explanation is in agreement with the results that our algorithm predicts many CRP binding sites for other genes in PCC7120 with statistical significance (Table S7 in Additional file 2).

Because the crp gene is widely distributed in many distantly related bacterial groups, including actinobacteria, aquificales, bacteroidetes, chlamydiae, chloroflexi, cyanobacteria, deinococci, firmicutes, fibrobacteres, planctomycetes, proteobacteria, spirochaetes, etc. (Figure S1 in Additional file 1), it was likely present in the common ancestor of the extant eubacterial lineages. Under this scenario, the crp gene in this ancestor was flexible enough to regulate different sets of genes. Alternatively, it is also possible that the crp gene originally evolved in a specific bacterial lineage and was subsequently spread to other groups via HGT. Such HGT events may benefit the recipient organism given the flexibility of CRP in regulating different biochemical activities, which are often coupled to environment stimuli leading to the generation of intracellular signaling molecule cAMP. Therefore, the crp gene in the ancestor of modern cyanobacteria was acquired by either vertical inheritance or an ancient HGT event from other lineages. Some cyanobacteria lost their crp genes since harboring the gene might not necessarily increase their fitness in their new environments. On the other hand, in other cyanobacteria, the crp genes were adapted to better meet their unique physiology and environmental requirements. In other groups of bacteria, crp evolved to regulate other lineage/species specific functions. For instance, it regulates the carbon and energy metabolisms in E. coli, cell-cell communication in Stenotrophomonas maltophilia 46, and quorum sensing in Vibrio cholerae 7, etc

Evolution of cis-regulatory binding sites is an interesting, but not well-studied problem. The 17 cyanobacterial genomes that do not encode a CRP orthologue provided us an excellent opportunity to examine the degradation of the CRP binding site in these genomes. It was expected that high scoring CRP binding sites do not present in those genomes, as previous studies have indicated that binding sites rapidly fade out when the corresponding TF was lost during the course of evolution 3247. We do see such fading in the CC9902, RCC307, WH7803, and WH8102 genomes (Figure S2 in Additional file 3). These observations were in consistent with the well-accepted rule that "if no such a TF in a genome, then no corresponding binding sites in the genome". However, surprisingly, in the rest 13 genomes, high scoring CRP-like binding sites seem to appear in the inter-TU regions with a higher probability than in the randomly selected coding regions (Figure S2 in Additional file 3), suggesting that there exist sequence patterns similar to CRP binding sites in these cyanobacterial genomes. A possible explanation of this unexpected observation would be that there exist in these genome TFs that recognize binding sites similar to CRP binding sites. Indeed, at least one member of the CRP/FNR superfamily is encoded in these 13 genomes, and it has been shown that the binding sites of the members of this TF superfamily are well conserved across many species with a wide range of evolutionary distance 40. Thus, the CRP-like binding sites found in these genomes are likely to be recognized by these non-CRP TFs of the superfamily. The specificity of these similar binding sites is likely to be achieved through the competition among homologous TFs for the same binding site, which is governed by their thermodynamic equilibrium 48.