Introduction

Gene expression patterns obtained by microarray experiments provide valuable information about gene-to-gene functional relationships, and thus they have been used to cluster functionally related genes since the dawn of microarray technology (Eisen et al. 1998). The use of microarray technology has rapidly spread and produced vast amounts of gene expression data for various species, and now these expression data are available in the public databases, such as NCBI GEO (Barrett et al. 2007), ArrayExpress (Rocca-Serra et al. 2003), TAIR (Swarbreck et al. 2008) and NASCArrays (Craigon et al. 2004). By analyzing these coexpression data, we can evaluate the similarity of expression patterns, and the gene pairs with similar expression patterns are called coexpressed gene pairs. Conceptually, gene coexpression can be defined with a small number of microarray experiments in a similar manner to traditional gene clustering, but “more is different”, as proposed by Anderson (1972). A large amount of expression data will yield a dramatically different scope of gene coexpression, and this gene coexpression information can be used as a fundamental gene map, rather than a simple gene classification.

In the past several years, coexpression approaches have been intensively applied to many biological targets, such as enzymes in a metabolic pathway, subunits of protein complexes and transcription factors (see reviews Aoki et al. 2007; Saito et al. 2008; Usadel et al. 2009). The coexpression data enable us to speculate about the functions of uncharacterized genes of interest and to search for new genes that are functionally related to a phenomenon under investigation. The success of gene coexpression approaches has given rise to several gene coexpression databases in the field of plant biology (Table 1).

Table 1 Gene coexpression databases
Full size table

We started to develop our coexpression database, ATTED (Obayashi et al. 2004), in November 2003. ATTED is one of the oldest gene coexpression databases, and it is presently available on-line as ATTED-II at http://atted.jp (Obayashi et al. 2007, 2009), with remarkable extensions from the original version. Comparisons of some of the details among these coexpression databases were provided in another recent review (Usadel et al. 2009). In this review, we would like to summarize a few successful examples of coexpression analyses with ATTED-II and to describe the most effective usages of ATTED-II based on the examples.

Brief introduction of ATTED-II

ATTED-II provides two different ways to examine gene coexpression information, a gene list view and a gene network view. The coexpressed gene list and the gene network for each gene in the Arabidopsis genome with the expression data were previously constructed, and the user can easily access the information. On the other hand, the coexpressed gene lists for multiple guide genes and the coexpression networks for the query genes are provided upon request, using the CoexSearch and NetworkDrawer tools, respectively. These are the two most popular tools in ATTED-II. The former is used for the “guide gene” approach, to find related genes with one or more guide genes, while the latter is used for the “narrow-down” approach, to analyze internal relationships among a set of genes and to identify the core genes in the set.

CoexSearch tool

The CoexSearch tool provides a list of genes that are coexpressed with the guide genes. Therefore, the guide genes are expected to have strong coexpression with each other, because this tool identifies the coexpressed genes based on the average coexpression strength for the guide genes. There is no strict criterion to judge whether the guide genes are strongly coexpressed, but the average values will not be meaningful when each of the guide genes is involved in different regulatory mechanisms.

NetworkDrawer tool

The NetworkDrawer tool accepts any set of genes and analyzes the internal relationships among the query genes. To draw the gene network from the lists of coexpressed genes, a threshold must be determined to define the coexpressed gene pairs. In ATTED-II, the three most strongly coexpressed genes for each gene are used to draw the network. This criterion was determined from the viewpoint of the user’s visibility. Greater numbers of genes can be incorporated into the network, and while the network may become more informative, it also tends to be more difficult to understand.

Quality of the coexpression data

In addition to its user-friendly interfaces, one of the most important features of ATTED-II is the continuous improvement of the coexpression data with the development of new calculation methods for gene coexpression (Obayashi and Kinoshita 2009; Kinoshita and Obayashi 2009). We have quantified the quality of the gene coexpression data by using Gene Ontology Annotation (Obayashi and Kinoshita 2009), and confirmed its improvement for every update (see CoexVersion for the version history; http://atted.jp/top_search.shtml#coexversion). The user can download all of the coexpression data for further analyses. Our coexpression data are actually used in several other databases and web tools, such as PRIMe (Akiyama et al. 2008) to produce network files, CoP (Ogata et al. 2009a, b) to find coexpression network modules, KaPPA-View3 (Tokimatsu et al. 2005) to integrate transcriptomics and metabolomics analyses on pathways, PosMed (Yoshida et al. 2009) for positional cloning, and Ondex (Lysenko et al. 2009) to integrate various omics data.

Other details of ATTED-II can be found in Obayashi et al. (2007, 2009).

Examples using ATTED-II with experimental verifications

To understand the strengths and weaknesses of the gene coexpression approach, we summarized successful studies that employed ATTED-II. The reports using the guide gene approach or CoexSearch tool are summarized in Table 2, and those using the narrow-down approach or NetworkDrawer tool are shown in Table 3.

Table 2 Examples using guide gene approach
Full size table
Table 3 Examples using narrow–down approach
Full size table

The studies by Ishihara et al. (2007), Takahashi et al. (2008) and Yamada et al. (2008) represent good examples of a single guide gene approach, where one target gene of interest for each study was already specified. However, this is not the most common scenario, and we often cannot decide which gene is the best one to use as the guide gene to identify functionally related genes. In such a case, multiple guide genes may be used to generate a single coexpressed gene list. The CoexSearch tool in ATTED-II provides a unified coexpressed gene list from multiple guide genes, by merging multiple gene lists based on the average MR (Mutual Rank) value, because our studies have shown that the MR value is more effective than the Pearson’s correlation coefficient (PCC), a popular coexpression measure (Obayashi and Kinoshita 2009). The actual MR values for each study are also shown in Table 2. In most of the successful studies, the genes with low MR values (i.e., tightly coexpressed genes) were used to design the experiments to verify the coexpression analyses, and the values ranged from MR = 1.4 in Ishihara et al. 2007 to MR = 67.7 in Takabayashi et al. 2009. However, there are two studies that used weaker coexpression (Bednarek et al. 2009; Sugano et al. 2010). We are not sure why these two cases were successful, but one possibility is that construction of a unified gene list with multiple guide genes may dramatically reduce the number of unrelated genes with relatively good average MR values. For example, in the study by Sugano et al. (2010), three guide genes (TMM, SDD, EPF1) were used to search for genes related to stomata development. If they had used a single guide gene, they would not have found STOMAGEN, the gene newly identified in their study, because this gene does not appear in the list of the top 300 coexpressed genes provided by ATTED-II. However, when they used three guide genes with the CoexSearch tool in ATTED-II, STOMAGEN appeared as the 17th ranked gene (Sugano et al. 2010). Bednarek et al. (2009) also used multiple guide genes, and they restricted their target to the cytochrome P450 gene family for the expected hydroxylation reaction in the glucosinolate biosynthetic pathway. This represents another approach to use weakly coexpressed genes as guide genes.

The narrow-down approach is useful to find the core coexpression module in a set of genes. Genes are often selected by using the pathway information in other databases, such as KaPPA-View (Tokimatsu et al. 2005), KEGG (Kanehisa et al. 2008) or Gene Ontology (GO, Ashburner et al. 2000). The NetworkDrawer tool in ATTED-II provides the internal structure of gene coexpression as a network, and the EdgeAnnotation tool provides it in a table. Since the output of the NetworkDrawer tool is a picture file and cannot be edited manually, we also provide the input format for other network drawers, such as Pajek (Batagelj and Mrvar 1998) or Cytoscape (Cline et al. 2007). Narrow-down approach was intensively used to identify genes for secondary metabolite pathways (Hirai et al. 2007; Yonekura-Sakakibara et al. 2007; Tohge et al. 2007; Sano et al. 2008; Okazaki et al. 2009; Sawada et al. 2009a, b). Okazaki et al. combined both the list approach and network approach. They first used the network approach to find the core coexpression module in a gene set for lipid metabolism, and then searched the coexpression data using the genes in the core coexpression module (Okazaki et al. 2009).

As described above, the MR and unified list methods are the keys for understanding the reason why the coexpression analyses worked well. In addition to these two reasons, the expression levels of the identified genes in Tables 2 and 3 were also investigated. As shown in Fig. 1, the identified genes were accumulated at the 50th to 80th percentile expression levels, and thus there may be some tendencies for highly expressed genes to be more suitable for coexpression analyses. The reason for this tendency is not straightforward, but one possibility is that the coexpression data of genes with low expression are less accurate, due to microarray noise, or that the phenotype of gene disruption appears more readily for highly expressed genes, as compared to those with lower expression.

Fig. 1
figure 1

Expression levels of the experimentally verified genes shown in Tables 2 and 3. Expression levels were determined using the average MAS5 value against all AtGenExpress developmental series experiments (Schmid et al. 2005). Following genes were used in this plot. As protein complex genes: PPL2, MCM2, MCM3, MCM4, MCM5, MCM7, NDF1, NDF2, NDF4, NDF6. As enzyme genes: UGT78D3, RHM1, CYP81F2, UGT89C1, OMT1, UGP3, MAM-IL, MAM-D. As TF genes: Myb28, Myb29, bHLHl15, LIM, C3H, HMG, hp_5g26. The other genes: PPL1, STOMAGEN, BASS5, NAI2. Two enzyme genes with low expression level are UGT78D3 and CYP81F2

Full size image

Suitable types of genes as targets of gene coexpression analyses

As shown in Tables 2 and 3, coexpression approaches have been applied to several experimental targets. Since these targets are quite diverse, we wondered whether there were any tendencies of the types of genes that are suitable targets of the gene coexpression analysis. To answer this question, we examined the coexpression tendencies in functional groups of genes with the same Gene Ontology (GO) term and with the coexpression data in ATTED-II version 5.5.

First, we evaluated background distribution by randomly selecting 20–30 genes and ploting the cumulative frequency distribution of their MR values among all pairs of the selected genes. Figure 2a shows the results for 200 repetitions. Most of the MR distributions were uniform, indicating that the MR values do not have strong biases in the ATTED-II coexpression data.

Fig. 2
figure 2

Degree of coexpression strength of Gene Ontology (GO) terms. MR is the coexpression measure used in ATTED-II. MR = 1 indicates the strongest coexpression. a Coexpression strength in randomly selected genes, b that in genes in each of 64 GO Biological Process (BP) annotations, c that in genes in each of 14 GO Cellular Component (CC) annotations, d that in genes in each of 34 GO Molecular Function (MF) annotations

Full size image

We then checked the MR distributions of the gene pairs in the gene groups sharing the same functional annotation and including 20–30 genes. The MR distributions for the genes for every GO Biological Process (BP) term are shown in Fig. 2b. Most of the GO BP groups are located in the upper-left area in this graph (Fig. 2b), indicating that the genes are coexpression modules in the gene set, because the preference of this region in the figure means the existence of strong gene coexpression, or a smaller MR value, among the selected genes. Figure 2c, d show the cases for GO Cellular Component (CC) and GO Molecular Function (MF), and the characteristics of these graphs were almost the same as those of GO BP.

To discuss these results quantitatively, the differences between each curve and diagonal line, or uniform distribution, in Fig. 2 were calculated, and the lists of GO terms with large deviations from the diagonal lines were generated, as potentially effective functional groups for gene coexpression analysis (Supplemental Table S1). From these lists, we chose some of the most, medium and least strongly coexpressed GO terms, as shown in Table 4. Basically almost all of the gene sets for the GO terms showed significant coexpression, as compared with the random distribution. It should be noted that statistical significance does not guarantee biological relevance, but it is possible that almost all of the gene pairs sharing the same GO annotation will have some weak but definite coexpression, as described later.

Table 4 The most, medium and least strongly coexpressed GO terms
Full size table

The most strongly coexpressed gene sets were the genes encoding the components of supramolecules, such as photosynthesis machinery, ribosomes and transcription machinery (Table 4). This observation is quite reasonable, because all of the components of a supramolecule should co-exist, to establish the three-dimensional structure of the molecule. In a similar manner, the genes encoding proteins that form tight complex structures will be strongly coexpressed.

The moderately coexpressed gene sets involved quite divergent genes, such as those for biosynthesis, stress responses or intercellular complexes (Table 4). The genes in this category have several strong coexpression cores and weak coexpression between the cores in the pathway.

Since almost all of the gene sets participate in coexpression modules, understanding the rarely coexpressed gene sets is important when using the coexpression data. The most rarely coexpressed gene sets were the genes for signaling pathways, such as transport and receptor binding (Table 4). Signaling pathways are mainly regulated at the protein-level, by modifications such as phosphorylation, whereas gene coexpression reflects the regulatory relationships of mRNA, and protein level regulation is not directly reflected. This limitation was clarified by Pitzschke and Hirt (2010) using the MAPK cascade as an example. The data in Table 4 agree with their statement, as in the case of the lower D values of signaling pathways. However, this result does not mean that the gene coexpression data will not be useful for studying signaling pathways, because the proteins regulating the signaling proteins, such as phosphatases and kinases, should also be regulated at the mRNA-level, since all of the participants in the signaling pathway must co-exist. When such weak coexpression information is used as in the signaling pathway, one of the important points to consider is to adopt other information about gene function at the same time, as in the case of Vicinanza et al. (2008). They discussed the coexpression network of phosphatases and kinases for the phosphoinositide system, using the human coexpression data provided in COXPRESdb, a coexpression database for animals (Obayashi et al. 2008). In their discussion, they employed a relatively low coexpression threshold (PCC > 0.4) and drew a gene network only for kinases and phosphatases (Vicinanza et al. 2008). The restriction of gene coexpression data by using the gene family is a promising approach, because it naturally combines genome information and coexpression information as in the guide gene approach. Although the identification of a signaling pathway is generally difficult, Sugano et al. (2010) recently reported a successful study about the identification of an intercellular signaling factor for stomata, STOMAGEN. In the case of transporters, which are also one of the most difficult targets, Sawada et al. (2009b) identified a transporter for glucosinolate biosynthesis.

How many genes should be examined?

Gene coexpression is defined by continuous values (MR or PCC), and there is no optimal threshold to define gene coexpression. In ATTED-II, we provide the top 300 MR genes for each guide gene. The reason why we selected 300 genes was based on a practical reason, since a gene list of this size can easily be checked visually. However, MR values around MR = 5,000 can still be meaningful to indicate the existence of gene relationships (Fig. 2), although such coexpression may also include indirect gene-to-gene relationships.

Based on the average distributions for each GO type in Fig. 2, we estimated the effective threshold to distinguish real coexpression pairs and random pairs. As a result, the most effective threshold was MR = 5,769 for BP, 3,625 for CC and 5,369 for MF. In ATTED-II, we provide the 300 most highly coexpressed genes for each gene, but it may be worthwhile to check for more weakly coexpressed genes, which can be obtained from the download page.

As one of the strategies to increase the reliability of coexpression data, ATTED-II provides a comparative view between Arabidopsis coexpression and rice coexpression, using orthologous genes. According to the determination of the significant MR threshold, we roughly set MR = 1,000, to highlight the conserved coexpression in other species.

Future directions

In this review, we have described the potential power of gene coexpression analyses by summarizing some successful examples with ATTED-II, and we also discussed a few limitations of these types of analyses. To overcome these limitations and enhance the power of ATTED-II, we are planning to integrate genome information and protein information into the coexpression information, in addition to continuously improving the coexpression data. The cis element analysis is one such approach to use genome information. ATTED-II also provides cis element prediction near the transcriptional start site, and the results were validated for some cases (Masuda and Fujita 2008). We are improving this cis element prediction function to understand gene coexpression. Although we did not discuss a comparison of coexpression among different species, the coexpression of orthologous genes in different species provides valuable information to enhance the reliability of coexpression data, and thus we will increase the number of target species in ATTED-II.