The cis-regulatory map of Shewanella genomes.

While hundreds of microbial genomes are sequenced, the challenge remains to define their cis-regulatory maps. Here, we present a comparative genomic analysis of the cis-regulatory map of Shewanella oneidensis, an important model organism for bioremediation because of its extraordinary abilities to use a wide variety of metals and organic molecules as electron acceptors in respiration. First, from the experimentally verified transcriptional regulatory networks of Escherichia coli, we inferred 24 DNA motifs that are conserved in S. oneidensis. We then applied a new comparative approach on five Shewanella genomes that allowed us to systematically identify 194 nonredundant palindromic DNA motifs and corresponding regulons in S. oneidensis. Sixty-four percent of the predicted motifs are conserved in at least three of the seven newly sequenced and distantly related Shewanella genomes. In total, we obtained 209 unique DNA motifs in S. oneidensis that cover 849 unique transcription units. Besides conservation in other genomes, 77 of these motifs are supported by at least one additional type of evidence, including matching to known transcription factor binding motifs and significant functional enrichment or expression coherence of the corresponding target genes. Using the same approach on a more focused gene set, 990 differentially expressed genes derived from published microarray data of S. oneidensis during exposure to metal ions, we identified 31 putative cis-regulatory motifs (16 with at least one type of additional supporting evidence) that are potentially involved in the process of metal reduction. The majority (18/31) of those motifs had been found in our whole-genome comparative approach, further demonstrating that such an approach is capable of uncovering a large fraction of the regulatory map of a genome even in the absence of experimental data. The integrated computational approach developed in this study provides a useful strategy to identify genome-wide cis-regulatory maps and a novel avenue to explore the regulatory pathways for particular biological processes in bacterial systems.

INTRODUCTION

Most of the genome-scale studies of bacterial regulatory systems have focused on the extensively studied model organism Escherichia coli. Despite the significant progress that has been made in identifying cis-regulatory elements in E. coli, 60–70% of the about 300 transcription factors (TFs) have not been characterized for their DNA-binding sites (1,2). Even less is known for the poorly studied microorganisms. As more bacterial genomes are being sequenced, identifying the cis-regulatory maps for these genomes remains challenging.

One common procedure for identifying putative cis-regulatory sites is to search for common motifs in the promoters of coregulated genes derived from microarray expression profiles or ChIP-chip experiments (3,4). A number of computational methods (5–8) using this strategy have been developed and successfully applied to identify cis-regulatory elements in both prokaryotic (9) and eukaryotic genomes (10,11). However, erroneously clustered ‘co-regulated’ genes often add noise to the step of searching for common motifs, and the limited amount of experimental data makes it difficult to extend these methods to the genome scale.

Over the past several years, several computational methods have been developed to identify DNA motifs on the genome scale in bacterial systems (12–16). Li et al. (12) presented an algorithm that extracts binding sites of regulatory proteins of a single genome without the knowledge of coregulated genes. This approach identified statistically significant dimer patterns and derived motif profiles from clusters of similar patterns. Using this method, they obtained about 160 distinct motifs from the E. coli genome. However, this approach had a low sensitivity: only one-third of the 60 characterized motifs of E. coli TF-binding sites (TFBS) were identified. Mwangi et al. (17) and Studholme et al. (18) performed similar analyses to predict DNA motifs in Bacillus subtilis and Streptomyces coelicolor, respectively. Using sequences from multiple genomes, the phylogenetic footprinting approach (19–21) provides an alternative way to identify regulatory elements of whole genomes without relying on experimental identification of coregulated genes. McCue et al. (22) applied whole-genome phylogenetic footprinting on E. coli and several γ-proteobacteria to find conserved motifs and employed a Bayesian clustering algorithm (13) to cluster motifs into distinct sets. Their study provided by far the most extensive collection of cis-regulatory elements in any bacterial genome. Similar strategies have also been applied on several other bacterial genomes (15,16,23,24) to find DNA regulatory motifs (see Supplementary Table 4). However, of the 181 motifs identified in E. coli using this method, 65% were shown to have only one or two target operons (http://www.people.fas.harvard.edu/~junliu/clust/). These results suggest that the coverage of the predicted cis-regulatory network is relatively low, considering that there are about 2700 transcription units (TUs) in E. coli (25). Several factors may contribute to affect the sensitivity and coverage of the phylogenetic approaches. First, species selection might not be optimal, as many E. coli DNA motifs may not be conserved in the other species. Second, algorithms for motif merging and clustering might be inefficient: conserved motifs initially identified by phylogenetic footprinting might be erroneously merged with other motifs in the clustering step, and some motifs in the final predictions might be redundant.

In this report, we present a new approach for genome-wide identification of cis-regulatory motifs in bacterial systems and its application in uncovering the cis-regulatory map of Shewanella oneidensis. We use two approaches to systematically identify DNA motifs and their target genes (regulons): one is to take advantage of the current abundant knowledge on the experimentally characterized transcriptional regulatory networks in E. coli to infer DNA motifs conserved in S. oneidensis; the other is to apply a new comparative genomics approach (Figure 1) that integrates phylogenetic footprinting, motif discovery with PhyloNet (21) and motif hierarchical clustering on multiple Shewanella genomes to predict novel DNA motifs on the genome scale. We assess our comparative approach by first examining whether the motifs identified using the five Shewanella genomes (S. oneidensis, S. denitrificans, S. frigidimarina, S. amazonensis and S. baltica OS155) are also conserved in the seven newly sequenced and distantly related Shewanella genomes (S. sp W3-18-1, S. putrefaciens, S. loihica, S. woodyi ATCC51908, S. sediminis HAW EB3, S. halifaxensis HAW EB4 and S. pealeana ATCC 700345). Furthermore, we compare our predicted motifs with known motifs and analyze the functional enrichment and expression coherence (EC) of the target genes of the predicted motifs. We find that 64% of the predicted motifs are conserved in the distantly related Shewanella species and 77 of these motifs are supported by at least one type of evidence. Applying our comparative approach on a focused gene set, the differentially expressed genes derived from microarray expression profiles of S. oneidensis during exposure to various metal ions, we were able to more sensitively identify regulatory motifs and their target genes potentially involved in metal reduction processes. Moreover, we find that our whole-genome comparative analysis is able to discover most of these motifs and their target genes, although in the absence of any experimental data.

Figure 1.

Flow chart of the procedure of identifying conserved cis-regulatory motifs in S. oneidensis by comparative analysis.

Shewanella oneidensis is a facultative, gram-negative γ-proteobacterium that can live in a wide variety of environments (26). Under anaerobic conditions, S. oneidensis can reduce various compounds, such as oxidized metals, inorganic chemicals and organic molecules (26–28). Its diverse respiratory capabilities ensure the great potential of S. oneidensis in bioremediation of both metal and organic pollutants. Although, many experimental studies are ongoing to determine the biological and biochemical characteristics of S. oneidensis, this organism is still far from being well understood. The complete genomic sequences of multiple Shewanella species provide valuable resources for DNA motif and regulon discovery using comparative genomics approaches. Identifying the cis-regulatory map in S. oneidensis will accelerate our understanding of the metabolism and gene regulation in this organism and ultimately facilitate its application in bioremediation.

MATERIALS AND METHODS

Datasets and genomic sequences

All complete genomic sequences used in this study were downloaded from the NCBI Genbank database. Shewanella oneidensis MR-1, S. denitrificans OS217, S. frigidimarina NCIMB 400, S. amazonensis SB2B, S. baltica OS155 and E. coli K12 were used for identifying cis-regulatory motifs in S. oneidensis. Eleven additional, recently sequenced Shewanella genomes (Shewanella sp. ANA-3, S. baltica OS185, S. loihica PV-4, Shewanella sp. MR-4, Shewanella sp. MR-7, S. putrefaciens CN-32, Shewanella sp. W3-18-1, S. woodyi ATCC51908, S. sediminis HAW EB3, S. halifaxensis HAW EB4 and S. pealeana ATCC 700345) were used to assess the conservation of the predicted motifs in these species. The datasets of experimentally characterized TF-target interactions in E. coli were downloaded from RegulonDB (version 5.0, released in 2006) (2). The series of publically available S. oneidensis microarray expression profiles under various conditions were collected from previous publications (28–30) and the M3D database (31) (see Supplementary Table 1 for downloading information).

Identification of orthologous genes between genomes

We identified orthologous genes by aligning all protein sequences from S. oneidensis (the anchor genome) to those from the other species using the NCBI BLASTP program (version 2.0) (6). Two genes were defined to be orthologous if all of the following three conditions are met: (i) their protein sequences are reciprocal best BLASTP hits between the two genomes; (ii) the BLASTP E-value is <1.0 × 10−10; and (iii) the BLASTP alignment covers ≥60% of the length of at least one sequence. Similar criteria have been used in a previous study (9).

Preparation of the promoter sequences of orthologous genes

Because genes in bacteria are organized in operons and most DNA motifs are located in the upstream regions of the first genes of the operons, we only retrieved the promoter sequence of each bacterial TU for motif search. In this study, a TU is defined as a set of consecutive genes in the same orientation whose intergenic distances do not exceed 40 nt (25). We collected the intergenic sequences up to 400 nt from upstream of the first gene of each TU in the anchor genome and its orthologs in the other genomes (not including any coding sequence if the length of the intergenic sequence is shorter than 400 nt). We refer to the promoter sequences of orthologous genes as ‘orthologous promoters’ in this article.

Inference of S. oneidensis TFBS from known E. coli regulatory interactions

We downloaded five datasets from RegulonDB (2) (version 5.0, 2006), including TF–target gene pairs, TFBS, promoters, gene products and alignment matrices, to compile the catalog of regulatory interactions for all experimentally characterized TFs in E. coli.

To use known regulons in E. coli to infer those in S. oneidensis, we first examined the conservation of E. coli TFs and their target genes (the first genes of regulated operons) in S. oneidensis by identifying orthologs between the two genomes. For each E. coli TF–target gene pair, we scanned the position specific weight matrix of the TF-binding motif along the promoter sequences of the orthologous target genes in the five Shewanella genomes (S. oneidensis, S. denitrificans, S. frigidimarina, S. amazonensis and S. baltica OS155) using the program Patser-v3b (32) to identify potential TFBS in these species. One standard deviation below the average of the Patser scores for all known E. coli binding sites of the TF was used as a cutoff. A binding site of an E. coli TF was considered conserved in S. oneidensis only if the promoters of the target gene orthologs in S. oneidensis and at least one of the other four Shewanella genomes had the binding sites whose Patser scores were not less than the cutoff.

Estimation of neutral distances between genomes

We estimated the neutral distances between all sixteen publicly available Shewanella species by examining the synonymous substitutions in coding sequences. The phylogenetic trees generated from the 16S rRNA sequences were used to guide the measurement of base substitutions.

There were 1670 genes conserved across all the sixteen Shewanella species. For each set of the orthologous genes, we first aligned their protein sequences using the program CLUSTALW (33), then converted the alignment of protein sequences into that of the DNA sequences, and finally calculated the synonymous substitution rate (Ks) using the program Codeml (34). Because the quality of multiple protein sequence alignments deteriorates as the sequence identity decreases (35), we only chose those alignments deemed reliable (protein sequence identity >75%) to estimate the Ks rates. Similar criteria have been used in previous studies (36,37).

Identification of novel DNA motifs on the genome scale in S. oneidensis

Identifying conserved DNA motifs in orthologous promoters by phylogenetic footprinting

As the majority of TFs in bacteria bind to DNA as dimers (38), we focused our investigation on the DNA motifs with palindromic patterns. We iteratively performed phylogenetic footprinting using the program CONSENSUS-v6c (32) to exhaustively identify all possible conserved palindromic motifs in each set of orthologous promoters. In each iteration, we searched for motifs with lengths of 17, 18, 22, 23, 27, 28, 32, 33, 37 and 38 nt. The identified significant sites (E < 10−10) were hen masked and the modified sequences were used for the next round of motif discovery until no more significant motifs could be found. All motifs were represented as position weight matrix profiles.

Finding the common motifs shared by multiple genes using PhyloNet

We used the core algorithm of PhyloNet (21) to identify motifs that are shared by multiple sets of orthologous promoters. PhyloNet is a BLAST-like algorithm for motif searching and comparing. It queries each motif profile obtained from the phylogenetic footprinting step against all other motif profiles to search for common motifs and groups the corresponding target genes. For each query, PhyloNet reports up to 10 significant motif profiles (P < 1 × 10−10) and their target gene sets.

Clustering similar motifs to generate nonredundant motifs

Because of exhaustive querying, PhyloNet may identify redundant motif profiles and target gene clusters. To remove the redundancy, we used hierarchical clustering algorithms to cluster similar motif profiles and merge corresponding target genes. The similarity of two motifs was measured by the average log likelihood ratio (ALLR) score of the global alignment of the two motifs (8) as defined below: where and are base frequencies at aligned positions, i and j, respectively, from two motif profiles; and are observed base counts at aligned positions, i and j, respectively; and p is the background base frequency. In general, the higher the ALLR score, the more similar the two motifs.

We implemented three hierarchical clustering methods, including the single-linkage, complete-linkage and average-linkage agglomerative algorithms, to cluster the motifs. Initially, every motif profile output by PhyloNet is the only element in each cluster. The clusters with similar motif profile element(s) are gradually merged together. In the single-linkage algorithm, the similarity of two clusters was measured by the ALLR score of the nearest elements between the two clusters. In the complete-linkage algorithm, the similarity was based upon the ALLR score of the farthest elements between the two clusters. In the average-linkage algorithm, the similarity was defined as the mean of ALLR scores of all possible pairwise comparisons of the elements between the two clusters. Two-motif clusters were merged if all the following conditions were met: (i) ALLR > 7.5; (ii) the aligned portion of two nearest (single-linkage algorithm) or farthest (complete-linkage algorithm) motif profiles between the two clusters was longer than 75% of any one of the motifs; (iii) the two clusters had at least two common target genes. These merging criteria were unlikely to merge dissimilar motif profiles based on assessment on motif profiles of known TFBS. The motif clusters kept merging the next closest cluster until no cluster pair met the merging criteria. Finally, we chose the motif profile with the most target genes to represent the whole cluster of motif profiles for further analysis. All unique target genes from different motif profiles within a cluster were combined as the target genes of the final motif profile.

Conservation of predicted DNA motifs in seven newly sequenced, distantly related Shewanella genomes

One assessment of our predictions considers the conservation of the predicted DNA motifs in the distantly related Shewanella genomes that were not used for motif identification. We collected all 16 Shewanella species whose genomic sequences are available, including the initial five species used for motif finding and 11 newly sequenced species, and calculated their pairwise neutral distances using the method described above. A Shewanella species is considered distantly related to the initial five species if its neutral distance to any of the five species and other selected distantly related species is not less than the minimum neutral distance between any two of the five species. Seven of the 11 newly sequenced species (S. sp W3-18-1, S. putrefaciens, S. loihica PV-4, S. woodyi ATCC51908, S. sediminis HAW EB3, S. halifaxensis HAW EB4 and S. pealeana ATCC 700345) meet such criteria and are used to examine motif conservation (see Supplementary Figure 2 and Supplementary Table 5). For each predicted motif, we scanned its position weight matrix along the promoters of all orthologous target genes in the seven distantly related Shewanella species using the program Patser-v3b (32) to identify the best matching sites. A motif is considered conserved in a distantly related Shewanella species only if the Patser score of its best matching site in this species is not less than that in S. oneidensis. As a control for spurious matches to low-specificity matrices, we scanned the same motif matrices in the upstream intergenic sequences of a set of randomly selected genes (excluding the orthologous target genes of the predicted motifs) from the seven Shewanella genomes.

Biological significance of predicted DNA motifs

Comparison of predicted DNA motifs to known TFBS

The binding motifs of a total of 30 known E. coli TFs that have orthologs in S. oneidensis were collected from RegulonDB (2). Focusing on palindromic motifs in this study, we compared our predicted motifs only to the 13 conserved E. coli TF-binding motifs with palindromic patterns. Motif similarities were measured by the ALLR statistics (8) as described above.

Functional enrichment analysis on the putative target genes of predicted motifs

As genes regulated by the same TF often have related functions in the cell, to find supporting evidence on our predictions, we examined the functional enrichment of target genes sharing the same motif in the anchor species (S. oneidensis) based on Gene Ontology (GO) (39) and KEGG pathway (KG) (40) annotations. The cumulative hyper-geometric distribution (11,41) was used to calculate the P-value (PGO/KG) of observing at least k genes having the same GO/KG functional annotation term among the n target genes of a motif profile. where M is the total number of genes with the particular GO/KG term and N is the total number of genes that have GO/KG annotations. Since bacterial genes are organized in TUs (operons), when calculating the GO/KG functional enrichment of the target genes, we considered all genes within the target TUs as long as they are annotated in the GO/KG databases.

Gene expression correlation analysis on the putative target genes of predicted motifs

Genes regulated by the same TF are expected to have correlated expression patterns in a series of microarray experiments. We used the EC score that was initially developed by Pilpel et al. (42) to measure the expression correlation of a set of target genes sharing the same motif. For each set of target genes, only the genes (K) that displayed varied expression across conditions were collected for study. The EC score of a set of K target genes is defined as p/P, where p is the number of gene pairs whose Euclidean distance between their mean and variance normalized expression profiles falls below a threshold, D (42); P = 0.5 × K × (K − 1), which is the total number of gene pairs in the set. We measured the significance of a resultant EC score by calculating the P-value (PEC), the chance that a cluster of randomly selected K-qualified genes from the entire expression data has an equal or greater EC score than the real cluster of target genes in 1000 permutations. In this study, we collected expression data from three series of microarray experiments on Shewanella, including exposure of Shewanella to different electron acceptors (28,29), cold shock (30) and H2O2 shock (31) (see Supplementary Table 1). Each of these microarray studies had 8–12 conditions, and the expression profiles of target genes under all the conditions were used to calculate the EC scores. We considered that a cluster of target genes had significant EC if its PEC <0.0183, which corresponds to a false discovery rate (FDR) of <0.05 (43), in at least one of these three microarray datasets.

Finding regulatory motifs and target genes involved in specific pathways from differentially expressed genes in microarray experiments

Differential gene expression in microarray experiments reflects cell's response to environmental changes or biological procedures. We can employ our comparative approach on orthologous promoters of differentially expressed genes derived from microarray experiments to identify regulatory motifs and their corresponding regulons that are potentially involved in specific biological processes. Instead of predicting cis-regulatory motifs individually from clustered genes that have similar expression patterns, our comparative approach is able to identify a set of putative regulatory motifs and corresponding target genes all at once. Moreover, working on a reduced motif search space, at a fixed significance level (the same P-value cutoff for PhyloNet to report motifs), this analysis should have a higher sensitivity in finding regulatory motifs than the whole genome analysis.

We applied this strategy on 990 differentially expressed genes derived from seven gene expression profiles of S. oneidensis during the exposure to a number of metal ions, including Fe(III), Mn(III), Mn(V), Cr(VI) and U(VI) (28,29), to identify regulatory motifs and corresponding target genes that are potentially involved in metal reduction.

RESULTS

Inferring regulatory interactions in S. oneidensis from verified transcriptional regulatory networks in E. coli

Shewanella oneidensis and E. coli both belong to the group of γ-proteobacteria and are assumed to share some common biology (26). We took advantage of the abundant knowledge on the E. coli genome and its regulatory networks and extrapolated it to the less characterized S. oneidensis genome.

We retrieved 62 E. coli TFs with known binding motifs and regulated genes (2) and examined the conservation of these TFs, their binding sites and target genes in S. oneidensis. We found that 30 of the 62 E. coli TFs have orthologs in S. oneidensis. Among the 450 TF–target pairs for these 30 TFs, 215 (48%) were conserved in S. oneidensis (Figure 2). We further examined whether the DNA-binding sites upstream of the E. coli target genes for the conserved TF–target pairs remain conserved in their counterparts in S. oneidensis. As shown in Figure 2, 24 of these 30 E. coli TFs had conserved DNA-binding sites in S. oneidensis. Of the 215 conserved TF–target gene pairs, 112 (52%) retained their TFBS in S. oneidensis. The list of 112 conserved regulatory sites and their positions in the S. oneidensis target genes are shown in Supplementary Table 2. These results suggest that ∼25% of the overall E. coli experimentally characterized transcriptional regulatory interactions are well conserved in S. oneidensis with regard to TFs, target genes and TFBS. The knowledge about the regulation of these E. coli regulons can be useful for inferring regulatory interactions in the genome of S. oneidensis.

Figure 2.

Conservation of the known E. coli transcriptional regulatory interactions in S. oneidensis. The logos of the E. coli TFBS were drawn based on the motif weight matrix models obtained from RegulonDB (Version 5.0). aThe number of binding sites used to build the motif profiles. Multiple binding sites may be from the same gene. bOnly the first genes of the target operons of a TF were considered when we counted the number of conserved target genes. A few target genes in RegulonDB were not included due to discrepancies in gene names.

Sequence divergence of Shewanella species

The use of multiple Shewanella genomes (detailed information is shown in Supplementary Table 3) for phylogenetic footprinting allows a high coverage of orthologous genes and facilitates the discovery of DNA motifs specific to this clade of species. As the ability of comparative sequence analysis to uncover DNA sites is intimately tied to the neutral rate of the genome sequence evolution (44), we estimated the neutral divergence along the lineage of the five Shewanella species for motif discovery to represent their evolutionary distances. On aligned coding sequences of orthologous genes from any two Shewanella species (34), we measured the synonymous base substitution rates (Ks) in terms of nucleotide substitution events per synonymous site since their last common ancestor. Figure 3 shows the phylogenetic tree of the five Shewanella species based on their neutral distances. The branches of this tree are measured by the medians of the Ks rates on alignments of all orthologous gene sets. Shewanella baltica OS155 is the closest species to S. oneidensis, with a divergence of about 0.65. This is almost the same as that between human and mouse, which have a divergence of about 0.64 (36,45,46). Shewanella dentrificans and S. frigidiamarina have similar distances of about 1.60 to S. oneidensis, more distant than that between Caenorhabditis elegans and C. briggsae, which have a divergence of 1.4 (36). Shewanella amazonensis is most distant to S. oneidensis, with a divergence of about 2.0, a distance slightly smaller than that between Drosophila melanogaster and D. pseudoobscura, which have a divergence of 2.4 (36). These species are sufficiently diverged that their orthologous intergenic regions are usually not globally well aligned, which means that conserved regulatory sites can be identified with low false positive and false negative rates using as few as three species (47).

Figure 3.

Estimation of neutral distances between five Shewanella species. The numbers shown are neutral substitution rates (Ks) measured by synonymous substitutions in coding sequences.

Identifying novel DNA motifs in the genome of S. oneidensis

Using S. oneidensis as the anchor species, we generated two sets of sequence data for identifying conserved DNA motifs that are potentially cis-regulatory elements. Dataset I contains promoter sequences of orthologous genes from both S. oneidensis and E. coli, along with orthologs from at least one of the other four Shewanella species. Including sequences from E. coli added phylogenetic divergence which helped in identifying the most conserved cis-regulatory elements and also facilitated the functional annotation of our predictions on the poorly characterized Shewanella genomes. Dataset I has 862 sets of orthologous promoter sequences. Dataset II contains promoter sequences of orthologous genes from S. oneidensis and at least two other Shewanella species. It has 1961 sets of promoter sequences, of which 80% have at least four orthologous sequences. Because of the closer evolutionary distances between the Shewanella species, Dataset II covers more orthologous genes than Dataset I. By subtracting predicted motifs common to both datasets from those predicted from Dataset II, we can obtain Shewanella lineage specific regulatory interactions, which will help us understand how the regulatory networks in the Shewanella species make them special for metal reduction and anaerobic respiration.

Our comparative analysis for genome-wide motif discovery includes three steps as shown in Figure 1. Table 1 lists the numbers of input orthologous promoter sets and the numbers of motifs identified at each step. First, we performed phylogenetic footprinting iteratively using the program CONSENSUS-v6c (32) to exhaustively identify significantly conserved palindromic DNA motifs with different lengths in every set of orthologous promoter sequences. In this step, we found 11 247 and 22 525 conserved motifs from Dataset I and II, respectively, with an average of 11–13 motifs for each promoter set. This procedure is unlikely to miss any significantly conserved motifs, but it identifies many similar motifs (with different lengths) from the same orthologous promoter set. In a regulatory network with a high degree of connectivity, each TF regulates multiple target genes, thus the cis-regulatory sites of the same TF can be found in multiple promoters. Hence, in the second step, we applied the core algorithm of PhyloNet (21) to identify common motifs shared by multiple promoters and cluster corresponding target genes. A large number of motifs that only occur in a single orthologous promoter set were discarded. After this step, the numbers of motifs in Dataset I and II were dramatically reduced to 203 and 1665, respectively (Table 1). However, the PhyloNet output motifs are still redundant, because PhyloNet is a BLAST-like algorithm for motif comparison and queries with different initial motif profiles may result in similar output motif profiles and target gene clusters. Therefore, in the third step, we further grouped similar motifs by hierarchical clustering methods. By assessing motif homogeneity within clusters and motif redundancy between clusters, we evaluated three agglomerative algorithms for motif clustering and found that the clustering results from the single-linkage algorithm were slightly better than those from the complete-linkage and average-linkage algorithms. Thus, we chose the motif profile clusters generated by the single-linkage algorithm as the final results in this article. In total, we identified 38 nonredundant motifs from Dataset I that covered 189 TUs, and 183 motifs from Dataset II that covered 824 TUs. The average numbers of target genes for the identified motifs were five and eight for Dataset I and II, respectively. Our predictions have a higher number of putative cis-regulatory interactions than those identified in previous studies in bacterial systems (12–18,22,23) (see Supplementary Table 4).

Table 1.

Numbers of motifs identified at each step of the comparative analysis for Dataset I and II

	Dataset I	Dataset II
The anchor genome(s)	S. oneidensis and E. coli	S. oneidensis
Other genomes	At least one other Shewanella species	At least two other Shewanella species
Number of sets of orthologous promoters	862	1961
Step I: Phylogenetic footprinting	11247	22525
Step II: PhyloNet searching	203	1665
Step III: Motif hierarchical clustering	38a (189)b	183 (824)

aNumber of nonredundant motifs ultimately identified.

bNumber of TUs covered by the predicted motifs.

We obtained fewer motifs from Dataset I than from Dataset II, mainly because many Shewanella genes do not have orthologs in E. coli and some motifs are not well conserved between E. coli and Shewanella species. The fact that 38 motifs were conserved across E. coli and Shewanella species suggests that these motifs may play common and crucial roles in gene regulation in both genomes. Actually, the majority of motifs obtained from Dataset I were also identified from Dataset II. After eliminating the motifs common to both datasets, we obtained 155 motifs unique to Dataset II, which are likely to represent regulatory elements that are specific to the Shewanella lineage (Supplementary Table 6).

Evaluating the biological significance of predicted regulatory motifs

We first examined whether the predicted DNA motifs in S. oneidensis obtained through comparative analysis on the five genomes, including S. oneidensis, S. denitrificans, S. frigidimarina, S. amazonensis and S. baltica OS155, were also conserved in other newly sequenced Shewanella species. Of the 11 recently sequenced Shewanella species, seven (S. sp W3-18-1, S. putrefaciens, S. loihica PV-4, S. woodyi ATCC51908, S. sediminis HAW EB3, S. halifaxensis HAW EB4 and S. pealeana ATCC 700345) were found distantly related to the initial five species used for motif discovery, based on their pairwise neutral distances and the criteria as described in Materials and methods section (Supplementary Figure 2 and Supplementary Table 5). These seven species were used to assess the conservation of the predicted motifs. Figure 4 shows that 64% of the 183 motifs identified from Dataset II were conserved in the orthologous promoters of at least three of the seven distantly related Shewanella species. In order to determine the background occurrence of the motifs, for each predicted motif, we checked its conservation in a control promoter set in which the orthologous promoters were replaced with the upstream intergenic sequences of the same number of randomly selected genes from the seven species. In only 2.2% of the control promoter sets was a motif conserved in at least three species. This suggests that the majority of our predicted motifs are evolutionarily constrained elements.

Figure 4.

Conservation of the 183 motifs identified from Dataset II in the distantly related Shewanella species. The black bars represent the percentage of motifs identified from the five Shewanella genomes that are also conserved in the seven newly sequenced Shewanella species. The white bars represent the results from the control sequence sets.

Because of the limited knowledge of TFBS in Shewanella, we chose to use the known E. coli TF-binding motifs to estimate the sensitivity of our predictions, which is defined as the fraction of known motifs that are correctly predicted. We used all 13 palindromic motifs from the 30 known binding motifs of E. coli TFs that have orthologs in S. oneidensis (2) to evaluate the sensitivity. Forty-six percent (6/13) and 69% (9/13) of the known motifs were correctly predicted from Dataset I and Dataset II, respectively (Supplementary Table 6). As shown in Figure 2, only a small fraction of the known E. coli transcriptional regulatory interactions were completely conserved in S. oneidensis. It is therefore not surprising that the prediction sensitivity for Dataset I was lower than that for Dataset II. Moreover, we compared the motifs identified from Dataset I with the predicted E. coli motifs that were computationally associated with TFs by Tan et al. (41). Four more motifs were supported by their study.

To further assess the biological significance of the predicted motifs, we analyzed the functional enrichment of their target genes (regulons) using the GO and KEGG pathway functional annotations and examined the EC of their target genes in various Shewanella microarray experiments. Of the total 183 motifs identified from Dataset II, the target genes of 24 motifs were significantly enriched for at least one biological function (PGO/KG < 1.0 × 10−5) and the target genes of 46 motifs displayed correlated expression (PEC < 0.0183 or FDR < 0.05) in at least one series of microarray experiments. Similarly, 6 and 13 of the 38 motifs from Dataset I were supported by the functional enrichment and EC of their target genes, respectively.

Figure 5 shows the numbers of the predicted motifs supported by the three types of evidence described above. Of the total 38 motifs identified from Dataset I, 17, 9 and 3 motifs were supported by at least one, two and three lines of evidence, respectively (Figure 5A and Table 2). Fifty-nine, 17 and 3 of the total 183 motifs identified from Dataset II were supported by at least one, two and three lines of evidence, respectively (Figure 5B and Table 3). The motifs with multiple lines of supporting evidence are the highest confidence predictions. Assuming that all the motifs with at least one type of supporting evidence are real motifs, we could estimate that the lower bounds of the specificities of our predictions are 45% (17/38) for Dataset I and 32% (59/183) for Dataset II, respectively. It is worth noting that even those motifs without any additional supporting evidence may be real, as 64% of them were conserved in at least three of the seven newly sequenced and distantly related Shewanella genomes.

Figure 5.

Venn diagrams showing the numbers of predicted motifs supported by different types of evidence, including matching to known TF-binding motifs, significant functional enrichment of the target genes in GO or KEGG pathway terms (GO/KG) and EC of the target genes in microarray experiments (EC). (A) Motifs identified from Dataset I (see Table 2 for detailed information). (B) Motifs identified from Dataset II (see Table 3 for detailed information).

Table 2.

The list of predicted motifs from Dataset I that have at least one type of supporting evidence

Motif Number	Motif consensus sequence	Known TF	Expression coherence of target genes	Biological functions in GO(G) or KEGG pathway (K) terms for which the target genes are enriched
1	aCTGTwtaTAtawACAGt	LexA	Yes	DNA repair (G)
2	ACgTcTAgAcGTCtA	MetJ	Yes	Methionine biosynthesis (G)
3	tTGATctagATCAa	FNR	Yes
4	yAaarNGCGCGCNyttTr		Yes	Structural constituent of ribosome (G), ribosome (G, K), protein biosynthesis (G)
5	ktAaAATkNcGCgNmATTtTam	CadCa	Yes	Protein biosynthesis (G), structural constituent of ribosome (G), ribosome (G, K)
6	AAtTtAAACgNNcGTTTaAaTT		Yes	Lipid catabolism (G, K)
7	TGTTGTaATATtACAACA		Yes	Pentose phosphate pathway (K)
8	aNTGaATtWWaATtCANt	ArgR
9	tGGTcWgACCa	FadR
10	tGCACcatwatgGTGCa	NtrC
11	WAaaAWycgCGcgrWTttTW		Yes
12	CACmAkATmTkGTG	YfeTa	Yes
13	aatgCgScGcatt		Yes
14	AtWTTgyatrcAAWaT		Yes
15	GCGtAtWaTaCGC	Nlpa	Yes
16	AAARggcgccwWwggcgccYTTT		Yes
17	aTtGGTaWtACCaAt	PdhRa

aMotifs that have been computationally associated with TFs by Tan et al. (41).

Table 3.

The list of predicted motifs from Dataset II that have at least one type of supporting evidence

Motif Number	Motif consensus sequence	Known TF	Expression coherence of target genes	Biological functions in GO (G) or KEGG pathway (K) terms for which the target genes are enriched
1	ACTGTaTatawatACAGT	LexA	Yes	DNA repair (G)
2	TaGACGTCTAgA	MetJ	Yes	Methionine biosynthesis (G, K)
3	tTGATctagATCAa	FNR	Yes	Oxidoreductase activity, acting on NADH or NADPH, quinone or similar compound as acceptor (G), thiamine metabolism (K), other energy metabolism (K)
4	CCGtWaCGG	CueR		Acetolactate synthase activity (G), butanoate metabolism (K)
5	tGCACcawwwtgGTGCa	NtrC		Starch and sucrose metabolism (K)
6	aATGatAAtNaTTatCATt	Fur	Yes
7	tGGTCWGACCa	FadR	Yes
8	aGCYWRGCt		Yes	Structural constituent of ribosome (G), ribosome (G, K), rRNA binding (G), protein biosynthesis (G)
9	RGCANNwWwNNTGCY		Yes	Motor activity (G), ciliary or flagellar motility (G, K), flagellum (K, G)
10	tAaAATKNcGCgNMATTtTa		Yes	Protein biosynthesis (G), structural constituent of ribosome (G), ribosome (K), rRNA binding (G)
11	tCGCGa		Yes	Oxidoreductase activity, acting on NADH or NADPH, quinone or similar compound as acceptor (G), other energy metabolism (K)
12	GCGSCGC		Yes	Structural constituent of ribosome (G), ribosome (G, K), protein biosynthesis (G)
13	AgcGAcKRYMgTCgcT		Yes	Cytochrome complex assembly (G), heme transporter activity (G), nitrogen metabolism (K)
14	GTAATWWWATTAC		Yes	Glucose metabolism (G), main pathways of carbohydrate metabolism (G), pentose phosphate pathway (K)
15	tTtAAACaNNtGTTT		Yes	Lipid catabolism (G, K)
16	RGACAaWtTGTCY		Yes	Unlocalized protein complex (G), ferredoxin hydrogenase activity (G)
17	GTWATATWAC		Yes	Pentose phosphate pathway (K)
18	TGTAaaNNWWNNttTACA	TyrR
19	TGaCANNatNNTGtCA	PhoB
20	CGTrATyACG			Oxidoreductase activity, acting on NADH or NADPH, quinone or similar compound as acceptor (G), other energy metabolism (K)
21	CTSSAG			Hydrogen-transporting ATPase, ATP synthase activity, rotational mechanism (G, K), proton-transporting two-sector ATPase complex (G)
22	GCGyATWATrCGC			Oxidoreductase activity, acting on NADH or NADPH, quinone or similar compound as acceptor (G), other energy metabolism (K)
23	KaATATATtM			Cytochrome complex assembly (G), heme transporter activity (G)
24	GATcagGTTaA			Thiamin biosynthesis (G, K)
25	ccTGATCAgg			Thiamin biosynthesis (G, K)
26	CWCcSSgGWG			Bacterial chemotaxis (K)
27	GNGcMCAYKWTAWMRTGKgCNC			Cell division (K)
28	TGGCgaAtaTtcGCCA			Type II secretion system (K)
29	GACATAWTATGTC		Yes
30	TATGSCATA		Yes
30	TATGSCATA		Yes
31	AcTTTACGTtaACGTAAAgT		Yes
32	aAAAagSssscNwNgssscctTTTt		Yes
33	gGCsATsGCc		Yes
34	aACNCSGNGTt		Yes
35	CNatRMTSAKYatNG		Yes
36	CCASTGG		Yes
37	AAAGTSACTTT		Yes
38	CMCCTWAGGKG		Yes
39	tAccYgAgTaaWttAcTcRggTa		Yes
40	ARAGYTARCTYT		Yes
41	CAaTWAtTG		Yes
42	RcATSATgY		Yes
43	aAaAttkGcgCmaaTtTt		Yes
44	CaAGGCCTtG		Yes
45	AAATctGCwGCagATTT		Yes
46	aacAtAAAGYNNRCTTTaTgtt		Yes
47	TAAYRTTA		Yes
48	TTtaTAcCTAGgTAtaAA		Yes
49	gCCaAcAmaGSCtkTgTtGGc		Yes
50	cgtWTGtTATAaCAWacg		Yes
51	cTTCGAAg		Yes
52	TTTTRYAAAA		Yes
53	WTGTSACAW		Yes
54	GcMcTATATAgKgC		Yes
55	gNcaaTGcTAgCAttgNc		Yes
56	AaaAARCGYTTttT		Yes
57	GGNAaAWTtTNCC		Yes
58	AtgGCSGCcaT		Yes
59	SCaTCawtGAtGS		Yes

By combining the 24 motifs inferred from known E. coli motifs, this study ultimately identified 209 unique cis-regulatory motifs in S. oneidensis. In addition to the motif conservation observed in the distantly related Shewanella genomes, 77 motifs (37%) were supported by at least one additional type of evidence. It should be pointed out that the numbers of motifs supported by EC and functional enrichment of target genes are underestimated, because many genes in S. oneidensis are unannotated and only a small number of expression datasets are available at this time.

Finding cis-regulatory motifs and target genes involved in metal reduction in S. oneidensis

Shewanella oneidensis is characterized by its diversified respiratory pathways, which ensure its capability of using a wide variety of metal ion substrates as terminal electron acceptors in anaerobic respiration (26,29). We collected seven gene expression profiles of S. oneidensis under the conditions of exposure to a range of metal ions, including Fe(III), Mn(III), Mn(V), Cr(VI) and U(VI) (28,29), and identified 990 genes (in 460 TUs) that displayed at least 3-fold change in transcript abundance in at least one of the microarray conditions. From the orthologous promoters of these TUs in S. oneidensis and the other four Shewanella species, we identified 31 nonredundant DNA motifs and corresponding regulons. Four of these motifs matched the binding motifs of E. coli TFs: Fnr, Fur, FadR and MetJ. Three of these four TFs have been previously shown to be involved in anaerobic metabolisms or growth under iron limitation conditions in bacteria. Specifically, Fnr is known to control a large number of genes necessary for anaerobic metabolisms and the transition from aerobic to anaerobic respiration (27), and Fur regulates expression of genes encoding iron storage proteins and iron-utilizing enzymes under iron limitation (48). In addition, recent evidence indicates that FadR governs fatty acid metabolism through interaction with the TF, ArcA, during anaerobic growth (49). The target genes of six predicted motifs were enriched for at least one GO or KEGG functional term, the target genes of 12 motifs had correlated expression, and five motifs were supported by both lines of evidence. In total, 16 motifs were supported by at least one type of evidence and 71% of them were conserved in the newly sequenced and distantly related Shewanella species (Table 4).

Table 4.

The list of predicted motifs potentially involved in metal reduction that have at least one type of supporting evidence

Motif Number	Motif consensus sequence	Known TF	Expression coherence of target genes	Biological functions in GO (G) or KEGG pathway (K) terms for which the target genes are enriched
1	TaGACGTCTAgA	MetJ	Yes
2	tGGTcWgACCa	FadR	Yes
3	AGCcTAgGCT		Yes	Structural constituent of ribosome (G), ribosome (G, K), protein biosynthesis (G), rRNA binding (G)
4	RGCANNwWwNNTGCY		Yes	Flagellum, flagellar assembly (G, K), bacterial motility proteins (K)
5	cGTCaAaWWtTtGACg		Yes	Ciliary or flagellar motility (G), flagellum, flagellar assembly (G, K), bacterial motility proteins (K)
6	AgcGActRYagTCgcT		Yes	Heme-transporting ATPase activity (G), cytochrome complex assembly (G), nitrogen metabolism (K)
7	YAAATgaNAacsgtTNtcATTTR	Fur
8	TGATctagATCA	FNR
9	gggRggCTWAGccYccc			Structural constituent of ribosome (G), ribosome (G, K), protein biosynthesis (G), rRNA binding (G)
10	CcatTGSCAatgG		Yes	Copper ion binding (G), heme binding (G)
11	WAaNcGCGCgNtTW		Yes
12	AaAGtTAaCTtT		Yes
13	TGtAAcANyrNTgTTaCA		Yes
14	aacAtAAAGYNNRCTTTaTgtt		Yes
15	TTtAAACaNNtGTTTaAA		Yes
16	aaAAarggNgcctNNgssscctTTTt			Signal transduction mechanisms (K)

We also identified several novel motifs that may be involved in regulating genes for metal reduction. First, the target genes of Motif 4 and Motif 5 shown in Table 4 were enriched for motor activity, flagellum and flagellar motility, and their expression profiles were highly correlated. Bacterial pili in both Geobacter sulfurreducens and S. oneidensis have been reported to serve as nanowires, transferring electrons from the cell surface to Fe (III) oxides (50–52). Whether the targets of these two motifs are involved in this process and how they function in metal reduction remain to be experimentally tested. Second, the target genes of Motif 3 and Motif 9 were enriched for ribosome, protein biosynthesis and rRNA binding. The change of expression of genes in these functional categories during the transition from aerobic growth to the anaerobic condition has been observed previously (53). Third, the target genes for Motif 6 were enriched for heme-transporting ATPase activity and cytochrome complex assembly that involves ATP-binding cassette transporters. Genome analysis of S. oneidensis indicates that this organism encodes 43 putative c-type cytochrome genes (26,51). Two c-type cytochromes, MtrC and OmcA (51), have been implicated in electron transfer during metal reduction. The targets for Motif 6 provide new candidates for this process. Fourth, the target genes of Motif 10 were another set of c-type cytochrome genes that were involved in copper binding and heme-binding. Finally, Motif 16 had a large set of targets (14 TUs) that were enriched in a signal transduction pathway. Characterization of these genes might provide insights into how signal transduction is involved in metal reduction.

Eighteen of the 31 motifs discovered in this set of differentially expressed genes were also identified in the whole genome analysis. This indicates that the comparative analysis across multiple species at the whole genome level is capable of identifying a large fraction of the regulatory map of the genome in the absence of any experimental data. But using experimental information to identify subsets of genes involved in specific processes can increase the sensitivity of motif detection and allow us to further refine the regulatory pathways for particular biological processes.

DISCUSSION

By inferring conserved known TF-binding motifs from the well studied model microorganism E. coli and by systematically searching for novel DNA motifs on the genome scale, we identified 209 motifs that cover 849 unique TUs in S. oneidensis. Sixty-four percent of the predicted motifs are conserved in the seven newly sequenced, distantly related Shewanella genomes and 37% of them are supported by at least one of the three additional lines of evidence, including matching to known TF-binding motifs and significant functional enrichment or EC of the target genes. In comparison with previous predictions (12–18,22,23) (see Supplementary Table 4), our study provides one of the most comprehensive cis-regulatory maps in bacterial systems with reasonable sensitivity and specificity. Moreover, by incorporating microarray profiling data into the analysis, our computational method also provides a novel strategy to characterize regulatory components for particular biological processes.

Inferring the regulatory networks from an extensively studied model organism is a direct way to identify conserved TFBS in a less characterized species. It also enables us to look into the evolution of the transcriptional regulatory networks across species. Several studies have been done on this topic (54,55). They are all based on the assumption that the regulatory interactions between a TF and its target genes in one species would occur in another species if both the TF and its target genes are conserved. This assumption might be true for those bacterial organisms that are evolutionarily close; however, it may not be appropriate for those evolutionarily distant genomes without taking into consideration the conservation of the TFBS. As shown in this study, for the 30 TFs that are conserved between E. coli and S. oneidensis, about 50% of the target genes in E. coli have orthologs in S. oneidensis, but only about 50% of these conserved targets bear the same TFBS. This observation implies that only about a quarter of the experimentally characterized transcriptional regulatory interactions in E. coli are still conserved in S. oneidensis. Hence, the previous estimate that about 40% of the E. coli transcriptional regulatory interactions are conserved in γ-proteobacteria (55) is likely an overestimate. Multiple factors can shape the bacterial transcriptional regulatory networks during evolution, such as gain or loss of TFs, target genes or TFBS, rearrangement of operon structures, and gene duplication through lateral gene transfer (56). Besides sharing the common targets, the orthologs of E. coli TFs in each Shewanella species may acquire new target genes to perform unique biological functions.

Species selection can significantly affect the effectiveness of comparative analysis in identifying DNA regulatory motifs in bacterial genomes. In order to cover more orthologous genes and discover regulatory elements specific to the Shewanella lineage, we chose five Shewanella genomes to identify DNA motifs genome wide. However, choosing evolutionarily closely related species can make it difficult to distinguish functional DNA-binding sites from background sequences, because the background sequences could be as conserved as the functional DNA sites. In this study, we used the genome-wide neutral mutation rates to estimate the evolutionary distances between bacterial species. Even though the five chosen Shewanella species are closely related across the bacterial phylogeny, they are sufficiently diverged so that three species are enough for identifying regulatory sites with low false positive and false negative rates (47,57). The result that the majority of the predicted motifs in S. oneidensis are also conserved in the seven newly sequenced, distantly related Shewanella genomes suggests that the predicted motifs evolved much slower than the background sequences. Therefore, it is likely that most of the predicted motifs are functionally constrained elements.

In addition to genome selection, several other factors can also significantly affect the quality of motif predictions. We applied an iterative search strategy for phylogenetic footprinting using the program CONSENSUS-v6c (32). This strategy allowed us to identify all possible significantly conserved motifs in each set of orthologous promoters and greatly enhanced the sensitivity of our predictions. However, many motif profiles identified from the phylogenetic footprinting step were highly redundant, and some of them might be false positives, appearing purely by chance in the orthologous promoters. Therefore, grouping similar motifs and eliminating false positives are essential to identifying true cis-regulatory motifs. As many false positive motifs are conserved in the orthologous promoters of a single gene and are not shared by other genes, PhyloNet (21) can efficiently eliminate a large number of false positive motifs by identifying common motifs shared by multiple promoters. With the subsequent step of motif merging using hierarchical clustering algorithms, we were able to further eliminate motif redundancy and generate a list of nonredundant motifs for the whole genome.

Our study provides a list of cis-regulatory motifs in Shewanella that are open for further validation. The detailed information of all predicted motifs is available online at http://ural.wustl.edu/databases.html. The fact that 64% of the predicted motifs are conserved in other distantly related Shewanella genomes and that at least 37% of the predicted cis-regulatory motifs are supported by additional evidence highlights the wealth of biologically significant information provided by our predictions. Furthermore, we have compared the motifs and target genes identified by the comparative genomics approach with those discovered from high-throughput microarray experiments in S. oneidensis. For example, from the microarray experiments on the S. oneidensis TF Fur and its mutant, Wan et al. (58) identified 21 coregulated target genes (operons) of Fur and predicted the consensus DNA motif of Fur binding sites from the promoters of these target genes using motif finding approaches. A motif profile identified by our comparative approach is very similar to that of the Fur binding motif obtained from the microarray data (Supplementary Figure 1A), and 11 of the 21 Fur target genes (52%) identified from microarray experiments were also shown in our predicted Fur regulon (Supplementary Figure 1B). Some of the Fur sites that we missed might be weak sites or specific to S. oneidensis, as they are not conserved in other species. Moreover, our approach predicted eight additional target genes that were not detected by the microarray experiments, and these genes are likely to be true targets, as they displayed strong EC with the 11 known target genes of Fur in a series of microarray experiments under stress conditions in which Fur is likely to be involved (59) (Supplementary Figure 1C). These results suggest that our comparative approach can extract useful biological information about regulatory networks in the absence of any experimental data, although having such data can provide a more informative view of the networks.

Although we expect that most of the predicted motifs are TFBS, we cannot rule out the possibility that some motifs may act as other functional elements. For example, some palindromic motifs could be part of the secondary structures of the RNA regulatory elements, such as riboswitches (60). However, considering the facts that the predicted motifs are all shared by multiple genes and that RNA regulatory elements usually are more conserved in secondary structure rather than in primary sequence (61), we think the number of RNA regulatory elements in our predicted motifs is very limited.

All predicted motifs and their target genes are putative components of the entire transcriptional regulatory networks. Using the nine motifs that match known TF-binding motifs, we can build the sub-networks containing TFs and their target genes. From such sub-networks, we can easily discover some useful biological information, such as the autoregulation of the TFs LexA, MetJ and NtrC, which has been reported in previous studies (62–64). With other sources of information, such as the distance constraint between TFs and their closest binding sites, phylogenetic correlation between TF and their regulons and binding specificity constraint for TFs (41), we may associate TFs with all predicted motifs to connect scattered components into a more complete transcriptional regulatory network that will allow us to gain insights into gene regulation in S. oneidensis.

We provide a strategy to identify bacterial lineage specific motifs by subtracting motifs conserved in more distantly related species (motifs from Dataset I) from all motifs identified in the lineage specific species (motifs from Dataset II). By this means, we predicted 155 Shewanella-lineage specific regulons, 43 of which are supported by additional evidence. Discovery of Shewanella-lineage specific regulatory interactions sheds light on the genomic differences that distinguish Shewanella from other γ-proteobacteria species and will help us understand how the distinct anaerobic respiration processes in Shewanella are regulated.

Instead of identifying cis-regulatory motifs individually from clusters of coregulated genes (11,42,65), our comparative genomics approach can predict a set of regulatory elements for particular biological processes all at once in the differentially expressed genes derived from microarray data. Our strategy can avoid problems that are often encountered in the former methods, such as limited expression conditions and incorrect or incomplete clustering of coregulated genes. With our approach, we successfully identified a set of known and novel cis-regulatory motifs involved in metal reduction in S. oneidensis. Those novel motifs can be high quality candidates for further experimental validation. Our predictions provide informative hints on which genes are possibly involved in the process of metal reduction and how they respond to environmental perturbations. These results show that utilizing experimental data can help elucidate not only regulatory motifs for specific biological processes, but also indicate that even in the absence of experimental data, comparative analysis using multiple related genomes can uncover a large fraction of the regulatory networks of an organism.

SUPPLEMENTARY DATA

Supplementary Data are available at NAR Online.

FUNDING

This work was supported by a grant from the US Department of energy (DE-FG02-05ER63972) to GDS.

Conflict of interest statement. None declared.