Plant Physiol. Illumina
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

First published online October 6, 2006; 10.1104/pp.106.085639

Plant Physiology 142:1589-1602 (2006)
© 2006 American Society of Plant Biologists

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Supplemental Data
Right arrow All Versions of this Article:
142/4/1589    most recent
pp.106.085639v1
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via CrossRef
Right arrow Citing Articles via ISI Web of Science (10)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Haberer, G.
Right arrow Articles by Mayer, K. F.X.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Haberer, G.
Right arrow Articles by Mayer, K. F.X.
Agricola
Right arrow Articles by Haberer, G.
Right arrow Articles by Mayer, K. F.X.
GENETICS, GENOMICS, AND MOLECULAR EVOLUTION

Large-Scale cis-Element Detection by Analysis of Correlated Expression and Sequence Conservation between Arabidopsis and Brassica oleracea1,[W]

Georg Haberer2, Michael T. Mader2, Peter Kosarev, Manuel Spannagl, Li Yang and Klaus F.X. Mayer*

Munich Information Center for Protein Sequences, Institute for Bioinformatics (G.H., P.K., M.S., L.Y., K.F.X.M.), and Institute of Stem Cell Research (M.T.M.), GSF National Research Center for Environment and Health, 85764 Neuherberg, Germany


   ABSTRACT
TOP
 ABSTRACT
RESULTS
 DISCUSSION
 MATERIALS AND METHODS
 LITERATURE CITED
 
The rapidly increasing amount of plant genomic sequences allows for the detection of cis-elements through comparative methods. In addition, large-scale gene expression data for Arabidopsis (Arabidopsis thaliana) have recently become available. Coexpression and evolutionarily conserved sequences are criteria widely used to identify shared cis-regulatory elements. In our study, we employ an integrated approach to combine two sources of information, coexpression and sequence conservation. Best-candidate orthologous promoter sequences were identified by a bidirectional best blast hit strategy in genome survey sequences from Brassica oleracea. The analysis of 779 microarrays from 81 different experiments provided detailed expression information for Arabidopsis genes coexpressed in multiple tissues and under various conditions and developmental stages. We discovered candidate transcription factor binding sites in 64% of the Arabidopsis genes analyzed. Among them, we detected experimentally verified binding sites and showed strong enrichment of shared cis-elements within functionally related genes. This study demonstrates the value of partially shotgun sequenced genomes and their combinatorial use with functional genomics data to address complex questions in comparative genomics.

Brassica oleracea enjoys a close evolutionary relationship to Arabidopsis (Arabidopsis thaliana). The two genera separated approximately 12 to 24 million years ago (Yang et al., 1999bGo). While Arabidopsis serves as a model for many research topics in plant genomics and plant biology, subspecies of Brassica oleracea cover a wide range of commercially important cultivars such as broccoli, cauliflower, and cabbage. The availability of the whole genome sequence for Arabidopsis as well as the large amount of genomic survey sequences (GSSs) for Brassica and their close evolutionary relationship make them useful for comparative plant genomic studies (Ayele et al., 2005Go; Katari et al., 2005Go).

Comparative genomics has been proven to be a powerful tool for the discovery of a large variety of functional elements by their conservation between related species. The usefulness of Brassica GSSs for the improvement of genome and specifically gene structure annotation in Arabidopsis, as well as for comparative studies of the repeat contents of both genomes, has been reported (Zhang and Wessler, 2004Go; Ayele et al., 2005Go; Katari et al., 2005Go).

In particular, it has been shown that comparative genomics approaches are able to detect genetic elements that are often difficult to discover due to their small size and/or limited information content. Examples include genetic elements like micro-RNAs and cis-regulatory elements (Wasserman et al., 2000Go; Cliften et al., 2003Go; Kellis et al., 2003Go; Jones-Rhoades and Bartel, 2004Go).

Comparative genomics detects cis-elements by their conservation between two or more evolutionary related sequences from orthologous genes. The assumption is that orthologs exhibit a common regulatory mode that is reflected in the conservation of transcription factor binding sites. Phylogenetic footprinting (Wasserman et al., 2000Go) and a related approach, phylogenetic shadowing (Boffelli et al., 2003Go), have been successfully applied over a wide range of genera ranging from bacteria to yeasts and mammals (McCue et al., 2001Go; Boffelli et al., 2003Go; Zhang and Gerstein, 2003Go). For plants, pioneering studies have been undertaken and encouraging results have been reported (Guo and Moose, 2003Go; Inada et al., 2003Go; Bao et al., 2004Go; Lee et al., 2005Go).

In addition to sequence conservation, a different popular approach uses functional information, mainly coexpression information, within one species to discover cis-elements. Powerful technologies to monitor transcriptional states and dynamics on a genome scale are well established and widely applied. The analysis of coexpressed genes under different conditions and states has been shown to be highly valuable for the analysis of shared cis-regulatory elements (Harmer et al., 2000Go). Diverse algorithms such as expectation-maximization or Gibbs-sampler have been adapted and applied to detect motifs that are overrepresented within sets of functionally related promoter sequences (Bailey and Elkan, 1994Go; Thijs et al., 2002Go; Tompa et al., 2005Go).

The majority of studies have used either sequence conservation or overrepresentation of motifs in promoters of coexpressed genes to discover cis-regulatory elements. However, some studies used a combination of both approaches to evaluate and/or screen detected motifs. For instance, Kellis and coworkers (Kellis et al., 2003Go) scored the motifs both by their overrepresentation within particular genomic regions and their conservation between four yeast (Saccharomyces cerevisiae) species to enrich for functional cis-regulatory elements. In a similar approach, conserved word occurrences have been further evaluated by expression data for a variety of species and kingdoms (Elemento and Tavazoie, 2004Go). In plants, no large-scale integrative studies for cis-element detection have been conducted so far.

Recent developments in the detection of cis regulatory elements integrate both phylogenetic information as well as coexpression information (Wang and Stormo, 2003Go; Moses et al., 2004Go; Sinha et al., 2004Go; Siddharthan et al., 2005Go). In our study, we apply PhyloCon to consider conservation between orthologous promoter regions as well as coexpression of genes within a species. Promoters of coexpressed genes and their respective orthologous counterparts are selected and subjected to a PhyloCon-based analysis. Initial profiles are generated from multiple alignments of orthologous promoter sequences. Motifs shared between different orthologous groups emerge by iteratively combining initial profiles. The high performance of PhyloCon has been demonstrated by the application on simulated data as well as on a test set of known yeast transcription factor binding sites (Wang and Stormo, 2003Go).

To integrate coexpression and sequence conservation for motif discovery, adequate information sources, coexpression and orthology information, are required. Recent contributions have provided both large-scale expression data for Arabidopsis and sequence data from Brassica. A large and high-quality expression dataset comprising about 800 microarrays has recently been made available (Craigon et al., 2004Go; Schmid et al., 2005Go). They monitor transcriptional states of the genome under various environmental conditions for different organs and tissues as well as during distinct developmental phases. These data allow us to identify genes that are coexpressed over a wide variety of different conditions. The high-quality Arabidopsis genome sequence and a large amount of Brassica GSSs provide a second information component. These sequences enable the search for conserved elements between the promoters of corresponding Arabidopsis and Brassica genes.

In this study, we undertook a comprehensive analysis of thousands of Brassica-Arabidopsis orthologous promoter pairs. To generate coexpression information, we analyzed a set of 81 microarray experiments from Arabidopsis totaling 779 chips. Promoters from coexpressed genes and their respective Brassica orthologous counterpart were selected. The resulting promoter sets were analyzed, and a large number of candidate sites have been discovered. These sites are derived from profiles, which are conserved between orthologous promoters and associated with coexpression. Many of the detected motifs are enriched for specific biological processes and pathways. Evaluation of our analysis with the aid of experimentally validated cis-regulatory elements from Arabidopsis confirms their significance. This study provides the basis for future cis-regulatory module analysis and analysis of regulatory circuits not restricted to Brassicaceae. It further demonstrates the benefits of partial genome sequences to address complex problems in comparative genomics.


   RESULTS
TOP
 ABSTRACT
 RESULTS
DISCUSSION
 MATERIALS AND METHODS
 LITERATURE CITED
 
The main goal of our study was the genome-wide discovery of candidate cis-regulatory regions in Arabidopsis. We selected PhyloCon to combine two sources of information for cis-element discovery: coexpression and sequence conservation. PhyloCon has been demonstrated to be very powerful both on biological and controlled artificial data (Wang and Stormo, 2003Go). Motif discovery by PhyloCon considers two independent axes of information, conservation of a motif between orthologous promoters and overrepresentation in a set of coexpressed genes. The necessary information was deduced from the identification of best candidates for orthologous upstream sequences between Arabidopsis and assemblies of Brassica GSSs. The Brassica genome has undergone a recent large-scale duplication event postdating the divergence of the two genera, Arabidopsis and Brassica (Town et al., 2006Go). Thus, a substantial number of Arabidopsis genes may not have a one-to-one ortholog in Brassica. However, previous studies have shown that recently duplicated genes are similar in their expression characteristics and cis-element composition (Blanc and Wolfe, 2004Go; Haberer et al., 2004Go). Thus, recently duplicated paralogs in Brassica can be expected to share high similarities with the corresponding Arabidopsis promoter.

To avoid potential misassignments, we applied a stringent bidirectional best BLASTN hit strategy to detect the best available candidates for orthologous promoter sequences (see "Materials and Methods"). In the following, we use the term orthologous promoters for these candidate pairs. Coexpressed Arabidopsis genes were determined using 779 microarray hybridization data (Craigon et al., 2004Go; Schmid et al., 2005Go). Figure 1 shows a workflow of the analysis undertaken.


Figure 1
View larger version (18K):
[in this window]
[in a new window]
[as a PowerPoint slide]
  		Figure 1. Analysis schema. Two sources of evidence, coexpression and conservation, were retrieved for a motif search in thousands of Arabidopsis upstream sequences. Expression data from 779 Affymetrix microarrays were normalized and an all-against-all correlation matrix was computed. For each gene, correlated genes form a CEG. Brassica GSSs were assembled and filtered for repeats. Brassica GSSs are depicted as green lines and Arabidopsis upstream sequences as blue lines. Orthologous sequences were derived by a bidirectional best BLAST hit strategy. All sequences of one CEG with associated orthology information represent a PAG. A PAG therefore consists of at least two OPG constituted by one Arabidopsis and one Brassica sequence. Each PAG was subjected to a motif search and all profiles, including intermediate profiles, were retrieved. Stepwise profile generation by PhyloCon was used to overcome potential shortcomings of the motif discovery procedure, in particular noisy data and partial coverage (see Fig. 3B). In an (hypothetical) example, PAG, a functional motif is shown as red box and it is present in three of four OPGs, while OPG 4 lacks it despite being coexpressed. A combination of OPG4 with any other of the three OPGs containing the motif will result in low quality profiles as marked by a question mark within the matrix. The motif in OPG3 is not covered by a (partial) GSS and any OPG combination may fail to detect the motif. However, the motif will be detected within the analysis of OPG1 and 2. As intermediate profiles were kept, the successive statistical evaluation identifies the profile as significantly overrepresented within the respective CEG while low quality profiles are excluded. In addition, its occurrence in OPG3 is uncovered by a genome-wide search for profile instances. Finally, candidate transcription factor binding sites are listed for each gene.

 

Figure 3
View larger version (28K):
[in this window]
[in a new window]
[as a PowerPoint slide]
  		Figure 3. Details of the analysis (A) gives an example how CEGs are defined for this analysis. Two (hypothetical) CEGs and their relationship are shown. Letters in circles stand for individual genes, lines indicate a significant correlation r ≥ 0.803, the 99% quantile. Anchor genes of the two CEGs are gene A and B, respectively. Note that the two CEGs share gene A, B, and C; however, there are genes (X and Y) that are significantly correlated to only one of the expression profiles of an anchor gene. This results in overlapping but not identical CEGs. B, Stepwise generation of matrices in a PhyloCon analysis of a particular PAG. Profiles are shown as multiple alignments of several OPGs, and each OPG consists of one Arabidopsis-Brassica promoter pair. Arabidopsis sequences are indicated by At, Brassica sequences by Bo. Alignments are transformed to PWMs (see "Materials and Methods") that are depicted here as sequence logos. Starting from aligned blocks of all pairwise OPG alignments, i.e. four promoter sequences (not shown), new matrices are generated by stepwise combinations of aligned blocks thereby refining borders and composition of the profiles. Intermediate matrices can be retrieved at each combining step.

 

Brassica Orthologous Upstream Sequences

We assembled a set of 5 67,365 Brassica GSSs by applying highly stringent clustering to minimize both redundancy within the GSSs and to prevent the generation of erroneous hybrid clusters by exclusion of repetitive and/or ambiguous sequences. Assembly and repeat masking/filtering resulted in 142,489 clusters with an average length of 987 bp totaling 140.6 Mb nonredundant sequences. The genome size of Brassica has been estimated to be about 600 Mb (Arumuganathan and Earle, 1991Go). Thus, the clustered and repeat-filtered sequences correspond to approximately one-quarter of the genome.

Next, we determined orthologous upstream sequences between Arabidopsis and Brassica by reciprocal BLASTN comparisons (E ≤ 10–10). Starting from 26,535 genes in MAtDB (Schoof et al., 2004Go), homologous sequences for 7,427 Arabidopsis upstream sequences were detected in the Brassica GSS assemblies. Applying the reciprocal best BLAST hit criterion, 4,007 were retained as putative orthologous sequence pairs.


Determination of Coexpressed Arabidopsis Genes

Expression data available from the Nottingham Arabidopsis Stock Centre (Craigon et al., 2004Go) and generated within AtGenExpress (Schmid et al., 2005Go) were used for the analysis. Affymetrix probes were remapped onto the current annotation of MatDB, and ambiguous probes were excluded from the analysis (see "Materials and Methods" for details). Quality filtering resulted in a total of 21,559 genes for which expression data were obtained. The expression experiments cover a broad variety of biological processes, tissues, and stages. Expression data were normalized and Pearson correlation coefficients (PCCs) of all against all genes were computed (see "Materials and Methods").

The background distribution was derived from the pairwise correlations of all 21,559 genes used in this study (Fig. 2 ). To define groups of coexpressed genes, the 99% quantile of this distribution was considered as significant (r = 0.803; Fig. 2). For each Arabidopsis gene, we assigned all genes exceeding a Pearson correlation of r ≥ 0.803 to its coexpression group (CEG). A particular CEG is defined by an anchor gene to which all other CEG members are significantly correlated (Fig. 3A ). A gene can be assigned to multiple CEGs, and therefore, two CEGs may share subsets of genes (Fig. 3A). This procedure enables us to differentiate genes and gene groups that participate in several distinct biological processes. From the initial 21,559 genes, 13,254 genes provided anchor points for CEGs, while 8,305 genes remained singletons. Approximately one-third (4,245) of the CEGs consisted of less than 25 members, and 49.4% (6,553) of all groups had less than 100 members, indicative of a stringent selection scheme (Supplemental Fig. S2). CEGs for each anchor gene are provided via a Web-accessible database hosting the results from this study (http://mips.gsf.de/proj/plant/webapp/expressionDB/index.jsp).


Figure 2
View larger version (8K):
[in this window]
[in a new window]
[as a PowerPoint slide]
  		Figure 2. Distribution of PCCs. A correlation matrix of all-against-all probe sets was computed and correlations, excluding self correlations, were binned into 100 bins of width 0.02. Bins are shown on the x axis and their relative frequency on the y axis. The dashed vertical line indicates the 99% quantile of the resulting distribution (r = 0.803).

 

Motif Discovery by PhyloCon

Each PhyloCon analysis group (PAG) is composed of the orthologous gene pairs of an individual CEG. Arabidopsis genes with no detectable orthologous upstream sequence in Brassica were removed from the analysis set. In the following, we refer to an orthologous pair of a Brassica GSS assembly and Arabidopsis promoter/upstream sequence as an orthologous promoter group (OPG). Thus, the collection of all OPGs of a particular CEG represents a PAG.

Elimination of genes without corresponding Brassica OPG significantly reduced the size and number of coexpressed groups, because an OPG has been identified on average for only about one-sixth of the Arabidopsis genes. In addition, each PAG had to consist of at least two OPGs. This filtering resulted in 4,540 PAGs that were subjected to a PhyloCon analysis.

PhyloCon initially creates profiles from pairwise OPG alignments. In subsequent cycles, merging and trimming profiles of preceding cycles generates derived profiles. Thus, profiles of later cycles are derived from alignments of an increasing number of distinct OPGs. An example for profile generation by PhyloCon is given in Figure 3B. For each PAG, both the final alignment matrices as well as profiles of previous cycles to which we refer as intermediate matrices were collected. This step reduces the likelihood to miss significant motifs in a noisy data set (see "Materials and Methods" for details). Analysis of all PAGs revealed a total of 322,079 preliminary profiles, including a large number of redundant intermediate matrices (see "Materials and Methods"). As (CT)n-repeats [or its respective complement, (GA)n] are very prominent in Arabidopsis promoters, we prefiltered consensus sequences of our matrices for the presence of such simple repeats. We analyzed the filtered matrices for overrepresentation within the associated CEG as compared to their frequency in all 21,559 analyzed Arabidopsis genes by testing against the cumulative binomial distribution. Within 3,861 PAGs, we detected at least one motif model that was significantly overrepresented for the respective CEG (P ≤ 0.01). Due to the partial coverage of the Brassica genome by GSSs, we could not identify putative orthologous promoter sequences for all members of a particular CEG. Hence, PAG sizes are generally smaller than the size of their corresponding CEG. Consequently, for only a subset of the CEG members, significant profiles are directly supported by conserved alignments. To overcome the limited PAG sizes and to transfer knowledge from the PAG analysis to all members of the CEG, we projected these profiles to all upstream sequences of the CEG. A profile was considered as a candidate motif for those genes of the CEG that contained at least one instance of the respective profile in their promoters.

Lengths of profiles predominantly range between 6 and 15 bp (Supplemental Fig. S1). This is in good agreement with sizes of known individual transcription factor binding sites. To estimate the number of candidate sites per gene, overlapping sites/instances of different motifs were merged (see "Materials and Methods"). Fusion of overlapping sites did not change the size distribution (Supplemental Fig. S1). This indicates that our profiles detect well-confined regions within the promoters. We found on average 7.3 nonredundant sites per gene and a total of 61,745 sites in 8,407 Arabidopsis genes (out of 13,254 genes contained in all CEGs). Each CEG and each gene can be queried for a list of significant profiles at http://mips.gsf.de/proj/plant/webapp/expressionDB/index.jsp. For the Web display, we assorted identical or nearly identical profiles exceeding a Pearson correlation of r ≥ 0.98 into one cluster (see "Materials and Methods"). However, as merged profiles generally result in a higher degeneracy compared to the single profiles, we only grouped but did not merge any profiles to retain the specificity for each profile for subsequent analysis.


Detected Profiles Match Known cis-Elements

We investigated whether profiles detected in this study match known cis-elements. For this purpose, we screened elements from the PLACE and AGRIS databases (Higo et al., 1999Go; Davuluri et al., 2003Go). Both databases contain experimentally validated cis-regulatory elements specific for plants (see "Materials and Methods").

In total, 537 motifs are contained within the two databases. However, there is a significant degree of redundancy both between the databases as well as within individual databases. In many cases, it is difficult to decide whether two motif variants constitute binding sites for two distinct transcription factors or represent two sites for one transcription factor. Therefore, we used all binding sites listed in both databases. For 255 sites (out of 537 sites; 47.5%), we detected a profile similar to the described motif within PLACE or AGRIS. Table I lists a selection of detected matches.


View this table:
[in this window]
[in a new window]

  		Table I. Known motifs in PLACE and AGRIS match discovered profiles

Left column shows sequence logos of profile discovered in this study, middle column lists name/identifier, and right column consensus sequence in PLACE/AGRIS for the motif matching the detected profile. This table represents a subset of motifs that match PLACE and AGRIS motifs.

 
Several reasons complicate the evaluation for profile matches to motifs reported in PLACE or AGRIS. First, motifs in PLACE are derived from various plant species. Hence, some motifs might not be present within Arabidopsis and Brassica. Second, many motifs are either reported as consensus sequences or experimental reports are restricted to only one specific site in a particular promoter. Particularly, the latter description is likely too specific as many transcription factor binding sites are degenerated. In addition, some consensus sequences do not describe binding sites for individual transcription factors but instead give the (degenerated) consensus for a family of transcription factors such as, for example, the Myb transcription factors (see Table I). This problem is especially pronounced for known motifs for which only a short core sequence is present and that are involved in the regulation of numerous pathways. Examples comprise the ACGT-element or the CAAT-box. Most importantly, the Brassica assembly only partially covers the Brassica genome, and the average length of the Brassica GSSs is about one-half the average length of Arabidopsis promoters used in this study. Thus, we are missing a considerable number of genes or promoter regions for comparison.

Nevertheless, our findings for several known motifs are consistent with experimental findings and functional enrichments described below. For instance, the PALBOX is frequently found in promoters of genes catalyzing steps in the phenylpropanoid biosynthesis. Consistently, we detected a significant enrichment of several profiles highly similar to it (e.g. Table II) in the flavonoid, phenylpropanoid, and lignin biosynthesis and in the category response to UV-C. A detailed description of the detected sites within the PAL promoter and their matches to known sites within this promoter is given in the last section. We also detected several profiles matching the G-box-related abscisic acid-responsive element GCCACGTG. In agreement with its regulatory function, these profiles were significantly overrepresented in the functional category response to abscisic acid stimulus (Table II).


View this table:
[in this window]
[in a new window]

  		Table II. Detected profiles enrich for GOSlim functional categories and KEGG biochemical pathways

Second column lists sequence logo representation of profiles discovered in this study that are statistically overrepresented in a functional category/KEGG pathway depicted in first column. Following columns show (1) the number of genes within the respective category that contain the motif in their promoters; (2) the number of genes in the genome annotated for the category; and (3) the Bonferroni-corrected probability that the enrichments is obtained by chance. Example results are shown for GOSlim annotations above, for KEGG pathways below the dashed line. The enriched motif for category "response to abscisic acid stimulus" has strong similarity to the known abscisic acid response element, the motif for category "phenylpropanoid biosynthesis" is similar to an overlapping P- and L-box present in phenylpropanoid biosynthesis genes (Fig. 5) and the motif for "trehalose biosynthesis" contains the core G-box.

 

Figure 5
View larger version (21K):
[in this window]
[in a new window]
[as a PowerPoint slide]
  		Figure 5. The C4H promoter analysis group. Figure 5A shows a (simplified) version of the core phenylpropanoid biosynthesis pathway leading to the synthesis of lignin monomers and anthocyanin pigments (not shown). Enzymes present in one PAG are written next to their reactions: PAL, 4CL, C4H, CCoAOMT (caffeoyl-CoA-0-methyltransferase). Figure 5B shows the upstream region of C4H, known motifs are displayed in bold text with the motif name under the site. Asterisks denote detected sites by PhyloCon. Note that P- and L-boxes partially overlap, that both boxes display a high degeneracy (P-box consensus: YTYYMMCMAMCMMC; L-box consensus: YCYYACCWACC) and are thus particularly difficult to detect.

 

Detected Profiles Are Overrepresented within Specific Functional Categories and Biochemical Pathways

Numerous studies demonstrated that coexpressed genes have an increased probability to be involved in a common biological process (DeRisi et al., 1997Go; Hughes et al., 2000Go; Schmid et al. 2005Go). Furthermore, correlations between the occurrence of particular motifs and specific functional categories have been demonstrated, for instance, in yeast (Kellis et al., 2003Go). To investigate whether we find enrichment for specific functional categories in the detected profiles, we made use of Gene Ontology (GO) information (GOSlim catalog) as well as information on biochemical pathways in Arabidopsis (KEGG pathways; Berardini et al., 2004Go). Both GOSlim annotations as well as the KEGG biochemical pathway information for Arabidopsis have been obtained from The Arabidopsis Information Resource (www.arabidopsis.org). GOSlim and KEGG assignments for the genes contained within our analysis set were selected and only categories that contain at least two genes were considered. Furthermore, we restricted our analysis for enrichment to GOSlim categories describing biological processes.

The enrichment in a particular KEGG pathway or a biological process defined by a GO term (see "Materials and Methods") was determined using the binomial coefficient against the genome-wide background distribution. P values have been Bonferroni corrected for multiple testing, and corrected P values ≤ 0.05 were considered significant (see "Materials and Methods"). For 538 out of 923 GOSlim categories and for 99 out of 292 KEGG pathways, we found significant enrichments for at least one candidate profile. Table II lists several example profiles and their respective functional or pathway category for which we detected a significant enrichment. To compile a list of candidate profiles for each gene, we employed a similar schema as applied for the analysis of profiles overrepresented in CEGs. Using profiles located within promoters of genes associated to pathway or functional category assignments, we analyzed for profiles that are significantly enriched in the respective category/pathway. This association rule even reports candidate motifs for genes that were not included in the analysis involving PAGs, e.g. singletons. One example is genes involved in the gibberellic acid (GA) biosynthetic pathway. For instance, a profile with a consensus sequence CACGTkTGGT (Table II) is found in 61.5% of the genes assigned to the GA biosynthesis pathway and hence is more than 4-fold enriched compared to a random expectation. Genes containing the motif in their promoters comprise almost all steps in the biosynthetic pathway for which Arabidopsis genes are known: an ent-copalyl diphosphate synthetase, the ent-kaurene oxidase GA3, several GA 20-oxidases, and a GA 3beta-hydroxylase forming the last step to biologically active GAs. Comprehensive results can be found and queried at the associated Web site (http://mips.gsf.de/proj/plant/webapp/expressionDB/index.jsp). Our findings and correlations will support the analysis of functional relations and aid to identify candidate regulatory sites within the promoter of individual genes of interest.


PhyloCon Sites Correspond to Experimentally Confirmed Sites

We evaluated our results using experimental findings. CRABS CLAWS (CRC), a member of the YABBY gene family, is required for nectary and carpel development in Arabidopsis (Lee et al., 2005Go). In our analysis, CRC was present in various CEGs. Their constituting members can group these into two distinct biological processes. In the first group, CRC coclusters with AP1, the floral organ identity gene AP3, and the SEPALLATA group (SEP1, SEP2, and SEP3). All these genes play critical roles in flower development, and AP3 and SEP1-3 have been suggested to be involved in CRC regulation (Lee et al., 2005Go). The second cluster group is enriched for genes involved in sporocyte morphogenesis and contains genes like MALE STERILITY 2 (MS2; Aarts et al., 1997Go), SPOROCYTELESS (SPL; Yang et al., 1999aGo), and several genes with their expression patterns indicative of a role within the sporocyte. The promoter of CRC has been shown to contain five distinct regions that are conserved between three different species (Arabidopsis, Lepidium africanum, and Brassica; Lee et al., 2005Go). Detailed promoter studies have shown that these five regions are required for correct CRC expression, and individual functional binding sites have been identified by site-directed mutagenesis. Within our analysis, we identified 22 different conserved sites. Nineteen sites were located within one or another of the five regions, and one site overlapped with an experimentally verified CArG-box (Fig. 4 ). It is noteworthy that several profiles were specific for one or the other of the biological process cluster groups. For example, a profile with the consensus sequence TGGAGGCA was present within the promoters of CRC and sporocyte-specific genes but missing in floral organ identity genes (Fig. 4). All sites detected in the CRC promoter are located within conserved regions.


Figure 4
View larger version (12K):
[in this window]
[in a new window]
[as a PowerPoint slide]
  		Figure 4. Sites within the promoter of CRC. At the top of the figure, a scheme of the CRC promoter is shown. Five previously reported enhancer regions (gray boxes A–E according to Lee et al., 2005Go) are depicted. Motifs detected in our analysis are displayed as colored boxes. Green, red, and cyan boxes designate motifs found in flower developmental genes. The respective genes are shown at the bottom left (AP1, AP3, SEP1, and SEP2), including At1g72290 that encodes an endopeptidase inhibitor cooccurring with these genes. Blue boxes are specific for genes involved in sporocyte formation (MS2, SPL) or expression patterns implying a possible role in this biological process (ATA1). Black boxes in the CRC promoter are profiles detected by PhyloCon but are either only sporadically present or completely absent within the shown coexpressed promoters.

 
Figure 5A depicts a PAG enriched for genes involved in phenylpropanoid biosynthesis. Within this group, an enzymatic chain involving phenyl-alanine-ammonia-lyase (PAL1, the entry point of the biosynthetic pathway), trans-cinnamate 4-monooxygenase (C4H/CYP73A5), p-coumarate 3-hydroxylase, and a caffeoyl-CoA O-methyltransferase-like protein is present. We found these genes to frequently cocluster within our analysis. In addition, many of these PAG clusters contained two isoforms of the coumarate CoA-ligase (4CL1, 4CL2) and a second variant of the PAL2. Tight coexpression of PAL1, C4H, and 4CLs has been reported not only for Arabidopsis (Mizutani et al., 1997Go) but has also been found in poplar (Populus spp.; Hertzberg et al., 2001Go). The PAL enzymes, C4H and 4CL1/2, make up the core of the phenylpropanoid pathway, while p-coumarate 3-hydroxylase and the methyl-transferase catalyze steps of a specific branch within the pathway, the synthesis of precursors of the lignin biosynthesis (Fig. 5A). For C4H, several motifs characteristic of enzymes of the phenylpropanoid pathway have been reported: four P-boxes, three A-boxes, and two L-boxes. As shown in Figure 5B, almost all box instances are either fully or partially covered by conserved sites detected from the analysis of several C4H containing PAGs.

In summary, the examples illustrate that conserved sites detected in this study correlate very well with known transcription factor binding sites.


   DISCUSSION
TOP
 ABSTRACT
 RESULTS
 DISCUSSION
MATERIALS AND METHODS
 LITERATURE CITED
 
The detection of cis-regulatory elements in higher eukaryotes is a major challenge in functional genomics. Bioinformatic sequence analysis of transcription factor binding sites is notoriously difficult, as cis-elements are difficult to distinguish from background. Two major approaches are commonly undertaken. In the first strategy, coexpressed genes are selected and analyzed for shared cis sequence elements. A second approach, phylogenetic footprinting, aims to detect candidate transcription factor binding sites from conserved regions in alignments of orthologous promoters. Both strategies have been demonstrated to be powerful (Zhang and Gerstein, 2003Go). Wang and Stormo (2003)Go proposed to enhance cis-element discovery by a combination of two sources of information, coexpression and sequence conservation.

In our analysis, we applied PhyloCon to discover cis-elements in Arabidopsis upstream sequences. Coexpression information for Arabidopsis genes was derived from a large set of microarray experiments (Craigon et al., 2004Go). For the orthologous reference genome Brassica oleracea, about 560-Mb redundant GSS were available (Ayele et al., 2005Go; Katari et al., 2005Go). After repeat masking and clustering, 142 Mb of nonredundant GSS were analyzed for orthologous Arabidopsis equivalents through bidirectional best hits. We identified 4,007 Arabidopsis genes with a corresponding best candidate orthologous promoter in Brassica. We detected a large number of candidate transcription factor binding profiles supported by sequence conservation and statistically significant overrepresentation within coexpressed genes. The mapping of these profiles on the 13,254 CEGs, i.e. including genes with no orthology information, revealed more than 60,000 sites in 8,407 Arabidopsis promoters (63.4% of all genes found in one or more CEGs).

One limitation in our approach is the incomplete Brassica genome sequence that covers approximately one-fourth of the genome. OPGs identified in this study represent the best available candidates for orthologous promoter pairs. Albeit PhyloCon uses sequence conservation for motif discovery, strict orthologous relationship of sequence pairs is not compulsory. For the identification of cis-regulatory elements, paralogous promoters that contain conserved regulatory regions are useful as well (Haberer et al., 2004Go). Further limitations also include genomic rearrangements (e.g. insertions, deletions, translocations, or duplications) that in the absence of positional information may affect the correct promoter identification for an OPG. However, noise potentially caused by these effects is limited by two provisions. PhyloCon generates new motifs by stepwise additions and combination of motifs found in individual OPG comparisons. Profiles from falsely assigned OPGs represent random matches unrelated to functional elements present in other OPGs of the respective PAG. Thus, they are unlikely to be added to the list of motifs reported. In addition, we statistically evaluated overrepresentation of both intermediate and final motifs in the respective group of coexpressed genes.

A number of complete plant genomes will be available in the near future and will help to circumvent some of the limitations encountered with partial genomes. Map information for these genomes will enable us to detect syntenic relationships and thus support the detection of true corresponding orthologous promoters. Map-derived synteny relations, however, are impaired by the highly dynamic nature of plant genomes. Genome, segmental, and tandem duplications are prevalent, and in plant genomes, gene families are often highly expanded (Arabidopsis Genome Initiative, 2000Go; Vision et al., 2000Go; Simillion et al., 2002Go; Bowers et al., 2003Go). For Brassica, a report describes a recent partial duplication of a 2-Mb contig (Town et al., 2006Go). This complicates the detection and definition of orthologous gene and promoter relationships. Specifically, one Arabidopsis gene may have none, one, or more ortholog(s) in Brassica, and a one-to-one mapping does not necessarily reflect the correct orthologous relationship.

Detection of cis-regulatory elements is known to be an error-prone process. Nevertheless, several observations indicate a successful enrichment for functional transcription factor binding sites in our study: sequence conservation of motifs between Arabidopsis and Brassica, enrichment of motifs in functional categories, and detection of known sites. Sequence conservation between evolutionary related species is generally considered as an indicator for either short divergence times or the functional importance of the respective elements. Insufficient sequence divergence imposes a severe problem for classic phylogenetic footprinting analysis based on sequence alignments as nonfunctional elements cannot be delimited from functional elements. From a large set of {lambda} clones, Windsor et al. (2006)Go identified orthologous upstream sequences between Arabidopsis and its close relative Boechera stricta. They reported significant sequence conservation within these sequences. Although Brassica is more distantly related to Arabidopsis than B. stricta, sequence conservation may still be caused by insufficient divergence times. Support for this assumption comes from a previous analysis of one of our examples, CRC. A phylogenetic footprinting analysis by promoter alignments identified five conserved regions between Arabidopsis and Brassica, each comprising up to several hundred basepairs (Lee et al., 2005Go). In contrast, in our study motifs discovered in the CRC promoter are more refined (Fig. 4). The analysis of recently diverged sequences in our analysis is far less likely to result in extended motifs compared to phylogenetic footprinting. Promoters between different OPGs of one PAG are functionally related by coexpression but not by evolutionary relationship. Thus, for evolutionary unrelated OPGs from one PAG, sequence similarity beyond shared cis-elements is expected to be equal to background promoter similarity. Delimited motifs consequently emerge from extended alignments after comparing profiles from different OPGs. This powerful feature of PhyloCon has already been demonstrated in a study of four closely related yeast species (Wang and Stormo, 2003Go). In our study, mean and median sizes of motifs detected are in good agreement with sizes of known transcription factor binding sites and indicate the detection of well-delineated elements.

We investigated for enrichments within functional categories by making use of GOSlim and the KEGG biochemical pathways annotations for the respective Arabidopsis genes. Many profiles detected are enriched in a wide range of biological functional categories involving metabolism (e.g. gluconeogenesis), development (flower development), signaling (abscisic and GA signaling) as well as cell maintenance tasks like ribosome biogenesis. Applying the guilt-by-association rule, the occurrence of particular profiles or the functional enrichment within particular CEGs may assist to transfer knowledge to genes of yet unknown functions. To assist in this task, we implemented a database and a Web portal providing structured access to all results of this study (http://mips.gsf.de/proj/plant/webapp/expressionDB/index.jsp).

We analyzed to what extent known Arabidopsis and plant cis-elements present within the AGRIS and PLACE databases overlap to the motifs detected within our analysis. We compared all cis-element entries present within the two databases with the profiles resulting from our analysis. We successfully detected 255 out of 537 elements present within the databases. Limiting factors in this analysis are the incomplete Brassica genome and the partial coverage of many Arabidopsis promoters by corresponding orthologous Brassica GSS contigs. An additional limitation is the partial coverage of the Arabidopsis transcriptome by Affymetrix GeneChips (21,559 out of 26,535 genes in MAtDB). Given these restrictions, the successful detection of 47.5% of described cis-elements from PLACE and AGRIS can be viewed as highly satisfactory. However, due to the incomplete data set as well as some limitations of motifs within the databases (e.g. single site reports, consensus sequences of transcription factor families, see "Results"), an exact global assessment of specificity and sensitivity for our results is not feasible.

Two examples we have studied in detail illustrate the correlation of sites detected in our analysis with regulatory elements involved in transcriptional regulation. CRC coclustered with several floral development genes that have been shown to interact with CRC (Lee et al., 2005Go). Five enhancers required for correct CRC expression have been reported (Lee et al., 2005Go). As shown in Figure 4, detected elements are preferentially located within these regions, and a CArG box confirmed by site-directed mutagenesis is covered by one of the cis-elements detected in our analysis. The C4H gene is involved in the phenylpropanoid pathway. Tight coexpression of PAL1, C4H, and 4CLs has been observed for both Arabidopsis and poplar (Mizutani et al., 1997Go; Hertzberg et al., 2001Go). Promoter elements for C4H have been characterized, and it has been shown that C4H contains four P-boxes, three A-boxes, and two L-boxes (Fig. 5B; Mizutani et al., 1997Go). Eight out of nine sites for the gene were identified. It is noteworthy that even the borders of several elements, a well-known problem in cis-element detection via phylogenetic alignments, were very well approximated.

In our work, we demonstrate that complex questions in comparative genomics can be addressed by using fragmented genome information and an integrative analytical approach, i.e. the combination of expression data with comparative sequence analysis and phylogenetic footprinting. Our analysis uses the comparison of a full and a partial genome sequence. The approach can be extended to additional partial or complete genomes to enhance the support for and the refinement of discovered motifs. The simultaneous analysis of several partial genomes, however, would decrease the number of OPGs available for the analysis, as best candidate orthologs have to be detected in multiple partial sequence sets. For instance, for the analysis of two genomes with coverage of one-quarter each, one would expect one-sixteenth of candidates on average to be present in both sets. Instead of a simultaneous analysis partial genomes may be sequentially subjected to a comparison against a complete genome. Subsequent processing and merging of results derived from pairwise comparisons would lead to a more comprehensive cis-element catalog.

Full genome sequences are labor and cost intensive, and high quality genome projects are expected in the near future for only a few model organisms and economically important species. Large-scale expression data will underlie similar limitations. Genome scale comparative genomics would thus have to rely on a few species that may be separated by large evolutionary distances, restricting the scope of comparative analyses. In plants, this problem is particularly accentuated as up to now only two genomes, rice (Oryza sativa) and Arabidopsis, have been analyzed extensively (Arabidopsis Genome Initiative, 2000Go; International Rice Genome Sequencing Project, 2005Go). On the other hand, shotgun GSSs can be produced at a fraction of the cost and provide a rapid and promising alternative for comparative genomics of closely related species. Besides being useful for gene discovery, these resources will be of importance for the detection of conserved genomic features beyond the genes and will allow the elucidation of additional functional elements like promoter elements and architectures.


   MATERIALS AND METHODS
TOP
 ABSTRACT
 RESULTS
 DISCUSSION
 MATERIALS AND METHODS
LITERATURE CITED
 

Brassica Dataset

Sequences were retrieved from the National Center for Biotechnology Information selecting for the keyword Brassica oleracea in the field Organism. The vast majority of the sequences represent GSSs of B. oleracea deposited by a sequencing consortium of The Institute for Genomic Research, the Cold Spring Harbor Laboratories, and Washington University. A total of 567,985 sequences (567,365 GSS) were obtained. The 567,365 GSS sequence reads represented approximately 384 Mb of sequence. A rigid clustering regime using the Harvester assembly pipeline (BIOMAX Informatics) was applied. The assembly method of Harvester is based on the CAP3 program (Huang and Madan, 1999Go). Default settings have been applied. This step also removed a large number of clones containing repetitive sequences. The remaining 190,513 GSSs defined 142,489 assemblies, with an average size of 987 bp, totaling 140.6 Mb of nonredundant genomic sequences of Brassica.


Determination of Candidate Orthologous Upstream Sequences

Individual Arabidopsis (Arabidopsis thaliana) upstream sequences were selected from the genomic sequence. Sequences were delimited either by the 5' neighboring gene or a maximum size of 3 kb (excluding the 5'-untranslated region [UTR]). Because 5' UTRs may harbor motifs or signals relevant to the transcriptional activity of a gene, 5'-UTR sequences were included in the analysis. To identify the best available candidates for orthologous promoter regions between partial genome information of Brassica and the complete Arabidopsis genome sequence, a bidirectional best Blast hit analysis strategy was applied. Upstream Arabidopsis sequences were compared against the Brassica GSS assemblies by BLASTN (E ≤ 1–10), while the Brassica GSSs were compared to the whole genome sequence of Arabidopsis. The genomic position of the highest scoring BLASTN hit of a Brassica GSS had to correspond to the location of the original upstream sequence. This modified reciprocal best hit criterion was used to group the candidate orthologous Arabidopsis and Brassica sequences.


Microarray Transcriptomics Datasets

Arabidopsis genome scale expression data have become available from a variety of microarray platforms. Among them are several cDNA arrays, both commercial and custom made, as well as two Affymetrix oligonucleotide GeneChips (http://www.affymetrix.com/products/arrays/index.affx?Arabidopsis). However, it is well known that comparisons among different platforms are problematic (The Toxicogenomics Research Consortium, 2005Go). Hence, we made use of measurements from a single platform only, the Affymetrix ATH1 GeneChip.

Experiments available from Nottingham Arabidopsis Stock Centre (http://nasc.nott.ac.uk, CD-ROM release as of November 2004; Craigon et al., 2004Go) and AtGenExpress (Schmid et al., 2005Go; kindly provided by Markus Schmid and Detlev Weigel) have been used in this study.


Mapping of the Affymetrix Probe Sets onto the Arabidopsis Genome

Due to annotation updates and enhanced gene modeling, GeneChip oligonucleotide mapping is frequently erroneous and outdated. Therefore, probe sets were recalculated using an enhanced oligonucleotide mapping against the Arabidopsis genome template.

All oligonucleotides present on the ATH1 GeneChip of Affymetrix (sequences downloaded from www.affymetrix.com as of October, 2004) were realigned against the coding sequence, and, for genes with associated full-length cDNA information, against the UTR sequences (MAtDB release from September 24, 2004; Schoof et al., 2004Go).

Oligonucleotides aligning to more than one gene and probes without perfect matches were excluded. For subsequent calculations, only probe sets with at least five probe pairs were considered. Most of the probe sets still consist of nine to 11 probe pairs. Four percent of the probes matched perfectly to at least two genes and led to partial unspecific estimates for 10% of the original probe sets, indicating the need for the realignments. We excluded those probe sets from our refined sets. In summary, expression measurements from 21,559 genes met the quality criteria and were used for subsequent analyses.


Statistical Processing

The statistical analysis of the expression data was carried out in R (R Development Core Team, 2004) using the FunDaMiner system (http://mips.gsf.de/proj/express). We calculated MAS 5.0, dChip (Li and Wong, 2001Go), and RMA (Bolstad et al., 2003Go) probe set summaries according to the redefined probe sets for every experiment. All probe set summary data were transformed to log-scale (basis 2). The complete dataset was normalized by applying the LMPN method. LMPN is based on the local polynomial regression fitting method loess (Cleveland, 1979Go; Cleveland et al., 1992Go) operating on MA scale (Dudoit et al., 2000Go). Nonlinear normalizations like the polynomial regression method are required if datasets consist of experiments from various researchers employing, e.g., different RNA extraction protocols like in our study. Ignoring this would introduce additional, technical rather than biological, correlative components. Replicates (usually around three) were summarized by the mean leading to more robust estimates of the real expression level. The correlation coefficients were computed for mismatch-corrected MAS 5.0 summaries.


Correlation Matrices and Distribution of Correlation Coefficients

For 779 measurements, i.e. microarray experiments, we computed the correlation matrix of all-against-all probe sets. The full matrix consists of about 4.65 x 108 (21,5592) correlation pairs. Correlations were determined as metric (Pearson) correlation coefficients. The full correlation matrix (except self correlations) served as background and the 99% quantile has been derived from this distribution (r = 0.803). Correlations with a correlation coefficient higher than the 99% quantile of the background distribution, analogous to a one-sided 1% significance level, were considered as relevant.


Extraction of Coexpressed Groups

For each of the 21,559 genes, its CEG was defined as those genes showing a Pearson correlation r ≥ 0.803. By this definition, one gene may be associated with multiple groups. By applying the above cutoff for correlations, we obtained 13,254 CEGs, while 8,305 genes remained as singletons.


PhyloCon Analysis

PhyloCon was downloaded from http://ural.wustl.edu/ approximately twang/Phylocon/ (Wang and Stormo, 2003Go).

A common problem in motif discovery is the degree of noise in the selected set of genes. One reason is that coexpression does not necessarily result from coregulation, as coexpression of two genes can be attributed to secondary effects (e.g. transcription factor cascades). Measurement errors, cross hybridization, biological variation, and erroneous annotations are additional sources of noise. In this study, the incomplete and fragmented genome of B. oleracea represents an additional difficulty. With an average length of about 1 kb for the GSS assemblies and about 1.8 kb for the Arabidopsis upstream sequences, orthologous information is available on average for only approximately 55% of the promoter region. Thus, even if all studied promoters of one specific CEG contain a conserved binding site, approximately one-half of the phylogenetic comparisons will on average fail to detect it, as alignments cannot cover the conserved site. However, PhyloCon computes candidate matrices in a stepwise manner. Starting from pairwise OPG alignments, it sequentially adds OPGs to the alignments from which new matrices are built. Importantly, PhyloCon allows for a report of intermediate matrices, i.e. matrices derived from preceding cycles.

To overcome the limitations of missing sites or noisy expression groups, we retrieved a maximum of 10 intermediate matrices per cycle. As additional parameters, we allowed for 200 temporary (or test) matrices per cycle and the number of SDs was set to 0.5 (for details, see Wang and Stormo, 2003Go). Both strands were analyzed.


Analysis of Profiles and Conserved Sites

Primary profiles were filtered for (CT)n- and (GA)n-repeats, as these repeats are very prominent on Arabidopsis promoters. Alignment matrices reported by PhyloCon were transformed into position weight matrices (PWMs) to generate a scoring function for sequence instances. For an alignment matrix of length m, we determined the number of occurrences nij of the four possible nucleotides i isin [A,C,G,T] in each column j = 1,2,,...,m. The following formula has been applied to transform this (4 x m)-count/frequency matrix into a (4 x m)-PWM (Hertz and Stormo, 1999Go):

Formula
where N is the number of instances in the alignment, pi the Arabidopsis background probability of nucleotide I, and nij the counts of nucleotide i at position j in the alignment of the instances found by PhyloCon. A single cell aij of a (4 x m)-PWM A is the respective score for nucleotide i at position j. Let S = s1,s2,..sm be a sequence of length m with each sj = 1,...,m representing a (particular) letter in the alphabet. A PWM A assigns a score to sequence S by summing up all j = 1,2,..m single cell scores aij, where i corresponds to letter sj. A sequence S is considered an instance of a PWM if its score exceeds a threshold score. This cutoff score is specific for each PWM. It is derived from the instance in the original PhyloCon alignment matrix with the lowest score. For each motif, we determined all occurrences in the Arabidopsis upstream sequences. Based on this mapping we derived expected and observed frequencies of motif occurrences.

For each PWM, we tested the statistical significance of its overrepresentation within the respective CEG in comparison to all 21,559 Arabidopsis upstream sequences (see "Materials and Methods"). P values were obtained from the cumulative binomial distribution and a PhyloCon PWM was considered to be significantly overrepresented for P ≤ 0.01.

To identify identical and almost identical profiles, profiles with a PCC of r ≥ 0.98 were grouped into clusters. Overlapping clusters were recursively merged. To determine the similarity between profiles, we used the PCC between two columns of a profile, as described in Schones et al. (2005)Go. To compare multiple columns of two profiles, the scores of each column comparison were summed up and normalized against the number of compared columns. For motifs of different lengths, we compared the shorter profile against longer profiles applying sliding windows. For these cases, the window with the highest PCC has been considered for the comparison.

Note that we did not merge any redundant profiles/PWMs, i.e. recompute a new alignment matrix derived from the merged profiles. Although this approach results in a significant redundancy, we avoid any flaws by low-quality matrices that potentially strongly alter the specificity of a merged profile. As a consequence, one would need to reassess findings for the merged matrices, e.g. enrichments of particular profiles in CEGs and functional categories, which have been obtained from the more specific individual profiles.

To derive an estimate for the number of sites detected in the PAG analysis, we merged detected instances/sites (not profiles) in each promoter if instances overlapped by more than 90%.


Known Motif Matches

To compare detected profiles with reported sites, all motifs listed in the AGRIS (http://Arabidopsis.med.ohio-state.edu/AtcisDB/bindingSiteContent.jsp) and PLACE (http://www.dna.affrc.go.jp/PLACE) databases were downloaded (Higo et al., 1999Go; Davuluri et al., 2003Go). PLACE and AGRIS derived motifs are given as consensus sequences, whereas profiles obtained during the course of our analysis are present as PWMs. Thus, a direct comparison was not feasible. Instead, we decomposed consensus sequences of known motifs into all exact sequences by replacing degenerated IUPAC letters into their respective bases. These sequences can then be tested for their probability to constitute an instance of a particular profile (see above). In analogy to the approach used for the determination of profile similarity, we applied sliding windows as described above.


Functional Categories

GOSlim annotation for Arabidopsis and the KEGG pathway map has been obtained from The Arabidopsis Information Resource (www.arabidopsis.org). All functional categories containing only one member have been excluded from subsequent analysis. Gene lists of categories were matched with the 21,559 genes used in this study. For each profile, we selected the genes containing the respective profile within their upstream sequences. Overrepresentation of the profile in a functional category was consequently checked for each GO annotation associated with the selected genes. P values for each test were obtained by cumulative binomial probability.

Formula
where n is the number of all studied genes associated with a specific GO annotation, x is the number of observed genes associated with this GO annotation and containing the profile, and p is the expected frequency of the profile, i.e. the number of promoters containing the profile divided by the number of all studied genes.

Profiles present in only one GO annotation were not considered (x > 1) as no reliable statistics can be computed for only one occurrence. Multiple testing corrections were performed by multiplication of the P value with the total number of assayed GO annotations for each profile.

For the KEGG pathways, we employed a similar binomial testing scheme. P values were corrected for multiple testing by the number of different KEGG pathways.


Supplemental Data

The following materials are available in the online version of this article.

Supplemental Figure S1. Size distribution of CEG derived profiles and sites.
Supplemental Figure S2. Size distribution of CEGs.


   ACKNOWLEDGMENTS
 
We thank Markus Schmid and Detlev Weigel for providing us microarray data from the AtGenExpress, and Chris D. Town from The Institute for Genomic Research for making the Brassica GSS dataset available to us prior to publication. The authors also wish to thank Louise Gregory for helpful discussions.

Received June 22, 2006; accepted September 24, 2006; published October 6, 2006.


   FOOTNOTES
 
1 This work was supported by the GABI program of the German Ministry of Education and Research (BMBF). Back

2 These authors contributed equally to the paper. Back

The author responsible for distribution of materials integral to the findings presented in this article in accordance with the policy described in the Instructions for Authors (www.plantphysiol.org) is: Klaus F.X. Mayer (kmayer{at}gsf.de).

[W] The online version of this article contains Web-only data. Back

www.plantphysiol.org/cgi/doi/10.1104/pp.106.085639

* Corresponding author; e-mail kmayer{at}gsf.de; fax 49–089–3187–3585.


   LITERATURE CITED
TOP
 ABSTRACT
 RESULTS
 DISCUSSION
 MATERIALS AND METHODS
 LITERATURE CITED
 
Aarts MG, Hodge R, Kalantidis K, Florack D, Wilson ZA, Mulligan BJ, Stiekema WJ, Scott R, Pereira A (1997) The Arabidopsis MALE STERILITY 2 protein shares similarity with reductases in elongation/condensation complexes. Plant J 12: 615–623[CrossRef][Web of Science][Medline]

Arabidopsis Genome Initiative (2000) Analysis of the genome sequence of the flowering plant Arabidopsis thaliana. Nature 408: 796–815[CrossRef][Medline]

Arumuganathan K, Earle ED (1991) Nuclear DNA content of some important plant species. Plant Mol Biol Rep 9: 208–218

Ayele M, Haas BJ, Kumar N, Wu H, Xiao Y, Van Aken S, Utterback TR, Wortman JR, White OW, Town CD (2005) Whole genome shotgun sequencing of Brassica oleracea and its application to gene discovery and annotation in Arabidopsis. Genome Res 15: 487–495[Abstract/Free Full Text]

Bailey TL, Elkan C (1994) Fitting a mixture model by expectation maximization to discover motifs in biopolymers. Proc Int Conf Intell Syst Mol Biol 2: 28–36[Medline]

Bao X, Franks RG, Levin JZ, Liu Z (2004) Repression of AGAMOUS by BELLRINGER in floral and inflorescence meristems. Plant Cell 16: 1478–1489[Abstract/Free Full Text]

Berardini TZ, Mundodi S, Reiser R, Huala E, Garcia-Hernandez M, Zhang P, Mueller LM, Yoon J, Doyle A, Lander G, et al (2004) Functional annotation of the Arabidopsis genome using controlled vocabularies. Plant Physiol 135: 1–11[Free Full Text]

Blanc G, Wolfe KH (2004) Functional divergence of duplicated genes formed by polyploidy during Arabidopsis evolution. Plant Cell 16: 1679–1691[Abstract/Free Full Text]

Boffelli D, McAuliffe J, Ovcharenko D, Lewis KD, Ovcharenko I, Pachter L, Rubin EM (2003) Phylogenetic shadowing of primate sequences to find functional regions of the human genome. Science 299: 1391–1394[Abstract/Free Full Text]

Bolstad BM, Irizarry RA, Astrand M, Speed TP (2003) A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics 19: 185–193[Abstract/Free Full Text]

Bowers JE, Chapman BA, Rong J, Paterson A (2003) Unravelling angiosperm evolution by phylogenetic analysis of chromosomal duplication events. Nature 422: 433–438[CrossRef][Medline]

Cleveland WS (1979) Robust locally weighted regression and smoothing scatterplots. J Am Stat Assoc 74: 829–836[CrossRef][Web of Science]

Cleveland WS, Grosse E, Shyu WM (1992) Local regression models. In JM Chambers, TJ Hastie, eds, Statistical Models in S. Wadsworth and Brooks, Pacific Grove, CA, pp 309–376

Cliften P, Sudarsanam P, Desikan A, Fulton L, Fulton B, Majors J, Waterston R, Cohen BA, Johnston M (2003) Finding functional features in Saccharomyces genomes by phylogenetic footprinting. Science 301: 71–76[Abstract/Free Full Text]

Craigon DJ, James N, Okyere J, Higgins J, Jotham J, May S (2004) NASCArrays: a repository for microarray data generated by NASC's transcriptomics service. Nucleic Acids Res 32: D575–D577[Abstract/Free Full Text]

Davuluri RV, Sun H, Palaniswamy SK, Matthews N, Molina C, Kurtz M, Grotewold E (2003) AGRIS: Arabidopsis gene regulatory information server, an information resource of Arabidopsis cis-regulatory elements and transcription factors. BMC Bioinformatics 4: 25–36[CrossRef][Medline]

DeRisi JL, Iyer VR, Brown PO (1997) Exploring the metabolic and genetic control of gene expression on a genomic scale. Science 278: 680–686[Abstract/Free Full Text]

Dudoit S, Yang YH, Callow MJ, Speed TP (2000) Statistical methods for identifying differentially expressed genes in replicated cDNA microarray experiments. Technical Report 578. Stanford University School of Medicine, Stanford, CA

Elemento O, Tavazoie S (2004) Fast and systematic genome-wide discovery of conserved regulatory elements using a non-alignment based approach. Genome Biol 6: R18

Guo H, Moose SP (2003) Conserved noncoding sequences among cultivated cereal genomes identify candidate regulatory sequence elements and patterns of promoter evolution. Plant Cell 15: 1143–1158[Abstract/Free Full Text]

Haberer G, Hindemitt T, Meyers BC, Mayer KF (2004) Transcriptional similarities, dissimilarities and conservation of cis-elements in duplicated genes of Arabidopsis. Plant Physiol 136: 3009–3022[Abstract/Free Full Text]

Harmer SL, Hogenesch JB, Straume M, Chang HS, Han B, Zhu T, Wang X, Kreps JA, Kay SA (2000) Orchestrated transcription of key pathways in Arabidopsis by the circadian clock. Science 290: 2110–2113[Abstract/Free Full Text]

Hertz GZ, Stormo GD (1999) Identifying DNA and protein patterns with statistically significant alignments of multiple sequences. Bioinformatics 15: 563–577[Abstract/Free Full Text]

Hertzberg M, Aspeborg H, Schrader J, Andersson A, Erlandsson R, Blomqvist K, Bhalerao R, Uhlen M, Teeri TT, Lundeberg J, et al (2001) A transcriptional roadmap to wood formation. Proc Natl Acad Sci USA 98: 14732–14737[Abstract/Free Full Text]

Higo K, Ugawa Y, Iwamoto M, Korenaga T (1999) Plant cis-acting regulatory DNA elements (PLACE) database. Nucleic Acids Res 27: 297–300[Abstract/Free Full Text]

Huang X, Madan A (1999) CAP3: A DNA sequence assembly program. Genome Res 9: 868–877[Abstract/Free Full Text]

Hughes TR, Marton MJ, Jones AR, Roberts CJ, Stoughton R, Armour CD, Bennett HA, Coffey E, Dai H, He YD, et al (2000) Functional discovery via a compendium of expression profiles. Cell 102: 109–126[CrossRef][Web of Science][Medline]

Inada DC, Bashir A, Lee C, Thomas BC, Ko C, Goff SA, Freeling M (2003) Conserved noncoding sequences in the grasses. Genome Res 13: 2030–2041[Abstract/Free Full Text]

International Rice Genome Sequencing Project (2005) The map-based sequence of the rice genome. Nature 436: 793–800[CrossRef][Medline]

Jones-Rhoades MW, Bartel DP (2004) Computational identification of plant microRNAs and their targets, including a stress-induced miRNA. Mol Cell 14: 787–799[CrossRef][Web of Science][Medline]

Katari MS, Balija V, Wilson RK, Martienssen RA, McCombie WR (2005) Comparing low coverage random shotgun sequence data from Brassica oleracea and Oryza sativa genome sequence for their ability to add to the annotation of Arabidopsis thaliana. Genome Res 15: 496–504[Abstract/Free Full Text]

Kellis M, Patterson N, Endrizzi M, Birren B, Lander ES (2003) Sequencing and comparison of yeast species to identify genes and regulatory elements. Nature 423: 241–254[CrossRef][Medline]

Lee JY, Baum SF, Alvarez J, Patel A, Chitwood DH, Bowman JL (2005) Activation of CRABS CLAW in the nectarines and carpels of Arabidopsis. Plant Cell 17: 25–36[Abstract/Free Full Text]

Li C, Wong WH (2001) Model-based analysis of oligonucleotide arrays: expression index computation and outlier detection. Proc Natl Acad Sci USA 98: 31–36[Abstract/Free Full Text]

McCue L, Thompson W, Carmack C, Ryan MP, Liu JS, Derbyshire V, Lawrence CE (2001) Phylogenetic footprinting of transcription factor binding sites in proteobacterial genomes. Nucleic Acids Res 29: 774–782[Abstract/Free Full Text]

Mizutani M, Ohta D, Sato R (1997) Isolation of a cDNA and a genomic clone encoding cinnamate 4-hydroxylase from Arabidopsis and its expression manner in planta. Plant Physiol 113: 755–763[Abstract]

Moses AM, Chiang DY, Eisen MB (2004) Phylogenetic motif detection by expectation-maximization on evolutionary mixtures. Pac Symp Biocomput 2004: 324–335

Schmid M, Davison TS, Henz SR, Pape UJ, Demar M, Vingron M, Schoelkopf B, Weigel D, Lohmann JU (2005) A gene expression map of Arabidopsis thaliana development. Nat Genet 37: 501–506[CrossRef][Web of Science][Medline]

Schones DE, Sumazin P, Zhang MQ (2005) Similarity of position frequency matrices for transcription factor binding sites. Bioinformatics 21: 307–313[Abstract/Free Full Text]

Schoof H, Ernst R, Nazarov V, Pfeifer L, Mewes HW, Mayer KF (2004) MIPS Arabidopsis thaliana database (MAtDB): an integrated biological knowledge resource for plant genomics. Nucleic Acids Res 32: D373–D376[Abstract/Free Full Text]

Siddharthan R, Siggia ED, van Nimwegen E (2005) PhyloGibbs: a Gibbs sampling motif finder that incorporates phylogeny. PloS Comput Biol 1: e67[CrossRef][Medline]

Simillion C, Vandepole K, Van Montagu MC, Zabeau M, Van de Peer Y (2002) The hidden duplication past of Arabidopsis thaliana. Proc Natl Acad Sci USA 99: 13627–13632[Abstract/Free Full Text]

Sinha S, Blanchette M, Tompa M (2004) PhyME: a probabilistic algorithm for finding motifs in sets of orthologous sequences. BMC Bioinformatics 5: 170[CrossRef][Medline]

The Toxicogenomics Research Consortium (2005) Standardizing global gene expression analysis between laboratories and across platforms. Nat Methods 2: 351–356[CrossRef][Web of Science][Medline]

Thijs G, Marchal K, Lescot M, Rombauts S, DeMoore B, Rouzé P, Moreau Y (2002) A Gibbs sampling method to detect over-represented motifs in upstream regions of coexpressed genes. J Comput Biol 9: 447–464[CrossRef][Web of Science][Medline]

Tompa M, Li N, Bailey TL, Church GM, De Moor B, Eskin E, Favorov AV, Frith MC, Fu Y, Kent WJ, et al (2005) Assessing computational tools for the discovery of transcription factor binding sites. Nat Biotechnol 23: 137–144[CrossRef][Web of Science][Medline]

Town CD, Cheung F, Maiti R, Crabtree J, Haas BJ, Wortman JR, Hine EE, Althoff R, Arbogast TS, Tallon LJ, et al (2006) Comparative genomics of Brassica oleracea and Arabidopsis thaliana reveal gene loss, fragmentation and dispersal after polyploidy. Plant Cell 18: 1348–1359[Abstract/Free Full Text]

Vision T, Brown DG, Tanksley SD (2000) The origins of genomic duplications in Arabidopsis. Science 290: 2114–2117[Abstract/Free Full Text]

Wang T, Stormo GD (2003) Combining phylogenetic data with co-regulated genes to identify regulatory motifs. Bioinformatics 19: 2369–2379[Abstract/Free Full Text]

Wasserman WW, Palumbo M, Thompson W, Fickett JW, Lawrence CE (2000) Human-mouse genome comparisons to locate regulatory sites. Nat Genet 26: 225–228[CrossRef][Web of Science][Medline]

Windsor AJ, Schranz ME, Formanova N, Gebauer-Jung S, Bishop JG, Schnabelrauch D, Kroymann J, Mitchell-Olds T (2006) Partial shotgun sequencing of the Boechra stricta genome reveals extensive microsynteny and promoter conservation with Arabidopsis. Plant Physiol 140: 1169–1182[Abstract/Free Full Text]

Yang WC, Ye D, Xu J, Sundaresan V (1999a) The SPOROCYTELESS gene of Arabidopsis is required for initiation of sporogenesis and encodes a novel nuclear protein. Genes Dev 13: 2108–2117[Abstract/Free Full Text]

Yang YW, Lai KN, Tai PY, Li WH (1999b) Rates of nucleotide substitution in angiosperm mitochondrial DNA sequences and dates of divergence between Brassica and other angiosperm lineages. J Mol Evol 48: 597–604[CrossRef][Web of Science][Medline]

Zhang X, Wessler SR (2004) Genome-wide comparative analysis of transposable elements in the related species Arabidopsis thaliana and Brassica oleracea. Proc Natl Acad Sci USA 101: 5589–5594[Abstract/Free Full Text]

Zhang Z, Gerstein M (2003) Of mice and men: phylogenetic footprinting aids the discovery of regulatory elements. J Biol 2: 11[CrossRef][Medline]




This article has been cited by other articles:


Home page
 Plant Physiol.Home page
K. Vandepoele, M. Quimbaya, T. Casneuf, L. De Veylder, and Y. Van de Peer
Unraveling Transcriptional Control in Arabidopsis Using cis-Regulatory Elements and Coexpression Networks
Plant Physiology, June 1, 2009; 150(2): 535 - 546.
[Abstract] [Full Text] [PDF]

Home page
 GeneticsHome page
Y. Wang, A. Diehl, F. Wu, J. Vrebalov, J. Giovannoni, A. Siepel, and S. D. Tanksley
Sequencing and Comparative Analysis of a Conserved Syntenic Segment in the Solanaceae
Genetics, September 1, 2008; 180(1): 391 - 408.
[Abstract] [Full Text] [PDF]

Home page
 Brief Funct Genomic ProteomicHome page
A. Krishnan and A. Pereira
Integrative approaches for mining transcriptional regulatory programs in Arabidopsis
Brief Funct Genomic Proteomic, July 16, 2008; (2008) eln035v1.
[Abstract] [Full Text] [PDF]

Home page
 Plant Physiol.Home page
K. Horan, C. Jang, J. Bailey-Serres, R. Mittler, C. Shelton, J. F. Harper, J.-K. Zhu, J. C. Cushman, M. Gollery, and T. Girke
Annotating Genes of Known and Unknown Function by Large-Scale Coexpression Analysis
Plant Physiology, May 1, 2008; 147(1): 41 - 57.
[Abstract] [Full Text] [PDF]

Home page
 Nucleic Acids ResHome page
H. W. Mewes, S. Dietmann, D. Frishman, R. Gregory, G. Mannhaupt, K. F. X. Mayer, M. Munsterkotter, A. Ruepp, M. Spannagl, V. Stumpflen, et al.
MIPS: analysis and annotation of genome information in 2007
Nucleic Acids Res., January 11, 2008; 36(suppl_1): D196 - D201.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Supplemental Data
Right arrow All Versions of this Article:
142/4/1589    most recent
pp.106.085639v1
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via CrossRef
Right arrow Citing Articles via ISI Web of Science (10)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Haberer, G.
Right arrow Articles by Mayer, K. F.X.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Haberer, G.
Right arrow Articles by Mayer, K. F.X.
Agricola
Right arrow Articles by Haberer, G.
Right arrow Articles by Mayer, K. F.X.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
ASPB Publications PLANT PHYSIOLOGY® THE PLANT CELL
Copyright © 2006 by the American Society of Plant Biologists