In order to determine how co-expression networks extend across metabolic processes, we selected three core metabolic pathways for analysis: fatty acid biosynthesis, starch metabolism, and mitochondrial leucine catabolism (Fig. 1). These pathways were selected in part because of the authors' extensive knowledge of their structure and biological implications (e.g., 33343536373839). Fatty acid biosynthesis is the process by which acetyl-CoA is used to generate acyl moieties that are used as components of cellular membranes, signaling molecules, and molecules that store energy (e.g., seed oils) 40. Starch is the principal polymeric storage form of glucose, and is stored either transiently in leaves or long-term in seeds 41. Leucine catabolism provides an alternate source of acetyl-CoA to sustain respiration and metabolic processes in the absence of photosynthesis 42. We assembled a list of genes with a demonstrated or putative function in these three metabolic pathways; these genes encode not only the enzymes of the "textbook pathway", but also enzymes important for synthesis of cofactors and transporters that are thought to be involved in these processes. Specifically, a set of 126 genes associated with these pathways was identified based upon one of three criteria: 1) genes that were demonstrated by experimental evidence to belong to one of these pathways; 2) genes assigned "putative" functionality in one of these pathways, based on sequence similarity and subcellular localization; or 3) genes encoding transporters and enzymes synthesizing co-factors required for these pathways [see Additional file 1].

Publicly available Affymetrix ATH1 chip-transcriptomic datasets comprising of 956 biological samples were used to infer patterns of transcript co-accumulation for the 22,746 Arabidopsis genes that are represented on this chip (see Methods section for details). These data are drawn from 72 experiments from NASCArrays 43 and PLEXdb 44 that represent a wide range of developmental, and environmental and genetic perturbations on Arabidopsis [see Additional file 2]. Data with poor reproducibility were discarded, and the data was normalized to the same mean and range (available at MetaOmGraph 11). We calculated the values of Pearson pair-wise correlation across these data for the set of 126 selected genes. The results were visualized as a graph, such that genes form the nodes, which are joined by an edge if the value of correlation between them is higher than selected threshold. Using Pearson correlation thresholds of 0.5, 0.6 and 0.7 yields three co-expression networks of 107, 77 and 62 genes, connected by 733, 444, and 204 edges, respectively (Fig. 2). The positioning of the nodes (genes) in these networks was produced by graph-layout software that places highly interconnected nodes close together 45. Thus, the absolute position of the node on the plot has no meaning, only its relative position versus other nodes: the most closely crowded nodes indicate genes with the highest co-expression. Within each of these co-expression networks the number of correlations between genes from the same metabolic pathway is significantly larger than in randomly generated networks with similar link structure (specifically, in the network at threshold 0.6, the proportion of within-pathway edges is 245/444, versus a mean of 112/444 in randomly rewired networks; this is equal to a difference of ~20 standard deviations; Fig. 3). As the correlation threshold is increased from 0.5 to 0.7, three distinct closely-interconnected clusters emerge, whose member genes closely correspond with the three metabolic pathways that were the focus of this study. We refer to the three closely-interconnected clusters identifiable in Fig. 2B as modules.

Starch is a complex molecule, whose metabolism comprises a network with many genes of known catalytic function but unclear physiological function 4647. Indeed, several genes may function both in the catabolism and biosynthesis of starch (e.g., pullulanase 48). At the 0.6 correlation level, the starch metabolism module contains 28 genes encoding all the known enzymes of starch metabolism. The module incorporates genes encoding proteins, for which, to date, there is no experimental evidence of involvement in starch metabolism, for example beta-amylase At2g32290 (BAM2), and three phosphoglucomutase-like proteins (At4g11570, PGM-like2; At1g70730, PGM-like3 and At1g70820, PGM-like4). On the other hand, alpha-glucosidase, an enzyme which historically had been considered to be involved in starch degradation, although recent experimental data from potato indicate otherwise 49, is not correlated with the Arabidopsis starch metabolic module (Fig. 2B). Although the processes of starch synthesis and degradation are temporarily separated over diurnal cycles in Arabidopsis leaves (reviewed in 50; 51), as is the transcription of many of the starch metabolic genes (5253; Foster and Wurtele, unpublished data), this is not reflected in the co-expression network of Arabidopsis; rather the two processes cluster together in a single module. Two factors in particular may contribute to starch synthesis and degradation transcripts being in a common module. First, Arabidopsis doesn't have a major starch-storage organ such as a tuber, thus an organ of Arabidopsis that is highly capable of starch synthesis may also be capable of a rapid shift to starch degradation, and the overall levels of all the relevant transcripts may be high. Indeed, this is the case for the diurnal fluctuations of the transcripts. Second, as several of the starch metabolic enzymes appear to function in both starch degradation and synthesis 48 (indeed the full extent to which this is true is not yet known), expression of such genes may be associated with both starch degradation and starch synthesis.

The fatty acid synthesis module contains genes encoding the complete set of enzymes required for the biosynthesis of 18-carbon fatty acids from acetyl-CoA: three nuclear-encoded subunits of acetyl-CoA carboxylase (CAC2, CAC3, BCCP1, BCCP2, and BCCP1-like), malonyl-CoA:ACP transacylase (MAT), acyl carrier protein (ACP1, ACP2&3 and ACP5), beta-ketoacyl-[ACP] synthases (KAS I, KAS II and KAS III), beta-ketoacyl-[ACP] reductase (KAR), beta-hydroxyacyl-[ACP] dehydratase (HD and HD-like), enoyl-ACP reductase (MOD1), stearoyl-[ACP] desaturase (FAB2 and FAB2-like2) and acyl-[ACP] thioesterases (FATA and FATA-like) (Fig. 4A). This module also contains all genes for the four subunits of plastidic pyruvate dehydrogenase, which produces acetyl-CoA (the precursor for fatty acid biosynthesis) from pyruvate. The apparent co-expression of the pyruvate dehydrogenase genes with fatty acid biosynthesis may indicate the direction of the major metabolic flux of acetyl-CoA, even though there are other fates of plastidic acetyl-CoA (e.g., biosynthesis of glucosinolates and amino acids, such as leucine and cysteine). Interestingly, the module contains genes for the final reaction in biotin biosynthesis 545556 and for lipoic acid biosynthesis, obligate cofactors of acetyl-CoA carboxylase and pyruvate dehydrogenase, respectively. In contrast, the acetyl-CoA synthetase (ACS) genes do not show any expression correlation, consistent with experimental evidence that this enzyme is not required to provide acetyl-CoA for fatty acid biosynthesis 335758.

The leucine degradation module contains genes for the four known enzymes of mitochondrial leucine catabolism (Fig. 5): branched-chain amino acid aminotransferase (BCAT), three out of four subunits of branched-chain alpha-keto acid dehydrogenase (BCDH), isovaleryl-CoA dehydrogenase (IVD) and methylcrotonyl-CoA carboxylase (MCC). Only a single BCAT is included in the module (BCAT2, At1g10070). This gene is not the one postulated by Schuster and Binder 59 to participate in leucine degradation. Consistent with these results, seven out of the eight genes in our module (except IVD) form a coherent subnetwork in the Arabidopsis gene network based on partial correlation 29.

To test whether these pathway-specific modules can be detected in the context of the expression of the entire genome, and to determine whether genes other than those encoding catalytic functions of the pathway are co-regulated, we calculated Pearson correlations between individual pathway-specific hub genes and the 22,746 genes represented on the Affymetrix chip. The genes used as hubs for each metabolic pathway are: At2g05990, encoding the enoyl-ACP reductase that is uniquely required in de novo fatty acid biosynthesis 60; At2g40840, starch metabolism hub gene coding for disproportionating enzyme 61; and the leucine catabolism hub gene, At4g34030, which encodes the β subunit of methylcrotonyl-CoA carboxylase 3942 (Table 1).

Despite the complexities associated with a genome of about 32,000 genes (TAIR7 Genome Release, April 23, 2007 62), and the challenge associated with the fact that the RNAs used for the microarray analyses came from plant samples that mix metabolically distinct cellular and tissue compartments, it was still possible to find expression correlations among genes of common metabolic pathways. Specifically, the majority of the genes that show the highest correlation with the fatty acid biosynthesis hub gene, the starch metabolism hub-gene, and the leucine catabolism hub-gene are the structural genes that code for the enzymes required in each of these metabolic pathways (Table 1). Similarly, for every gene in the module, we found genes correlated with it (above the set threshold) among 22,746 genes on ATH1 array and checked how many of those correlated genes are also in the same module ("in" links) and how many are not in module ("out" links). We found that at 0.7 threshold, for 19 genes out of 22 in the fatty acid biosynthesis module, the number of "in" links is higher than the number of "out" links, and the same is true for 7 out of 8 leucine catabolism genes, and 9 out of 16 starch metabolism genes at the threshold of 0.8. Thus, these genes are preferentially interconnected among themselves with fewer connections to other genes, which defines them as a module in the context of the Arabidopsis genome.

Analysis of co-expression also provides means of querying the consequence of genome duplications. As with other plant genomes, Arabidopsis has undergone whole-genome duplications (polyploidization), followed by radical gene silencing, gene turnover, and expansion of gene families, the most recent of which is thought to have occurred about 25 million years ago 6364. In our analysis, At5g52920, one of the fourteen Arabidopsis genes that code for pyruvate kinase, is highly correlated with the fatty acid biosynthesis hub-gene (Table 1). Pyruvate kinase catalyzes the conversion of phosphoenolpyruvate to pyruvate, the immediate precursor of the acetyl-CoA used for fatty acid biosynthesis 65. These data, therefore, help to constrain the complexity associated with the polyploidization, and indicate that of the fourteen pyruvate kinase genes that occur in the genome, At5g52920 is most likely the one predominantly utilized for fatty acid biosynthesis.

In addition, the co-expression of pyruvate dehydrogenase and pyruvate kinase with the fatty acid biosynthesis genes implies that the functional pathway commonly thought of as "fatty acid synthesis" (acetyl-CoA to fatty acids) could be expanded to begin with phosphoenolopyruvate (PEP), the substrate for pyruvate kinase (Fig. 4A).

A set of potential regulatory genes whose expression patterns highly correlate with each of the pathway-specific hub-gene was identified in this analysis (genes marked with an asterisk, Table 1). Although these correlations cannot be extrapolated to imply a causative role, identification of candidate regulatory genes delineates hypotheses and suggests experimental approaches for identifying the regulatory function of these genes.

One such putative regulatory gene encodes a protein tyrosine phosphatase/kinase gene SEX4 (At3g52180) (Table 1 and Fig. 6A). Its presence in this module would lead us to hypothesize an involvement in regulation of starch metabolism. Indeed, this gene was recently shown experimentally to have a regulatory role in starch metabolism; knockout alleles of At3g52180 have a starch excess phenotype 32. Similarly, a gene encoding a putative protein kinase (At1g26150, AtPERK10 66) is correlated with genes in the starch metabolism module. We directly checked whether this gene might play a role in regulation of starch synthesis.

We obtained three independent alleles, named Atperk10-1, Atperk10-2, and Atperk10-3, corresponding to T-DNA insertions at gene AtPERK10 locus. Homozygous mutant lines stained for starch using Lugol solution showed an increase in starch accumulation (data not shown). Figure 7 shows a quantitative comparison of starch content in wild type and Atperk10-3 homozygous allele.

Interestingly, the starch hub-gene shows a high negative correlation with two Rab-like, GTPase-coding genes (At5g47200; RAB1A and At4g17530; RAB1C; Fig. 6B). In eukaryotes, proteins from this family of small GTP-ases are regulatory and act as signaling molecules in diverse processes 67. The Arabidopsis protein that is phylogenetically the closest homolog to Rab-like GTPase is RAB1B (At1g02130), which is localized to the ER, and as in mammalian systems, it is important for Golgi apparatus development 68. RAB1A and RAB1C are predicted to be secretory and might have function similar to that of RAB1B, or some novel plant-specific regulatory function, which appears to be negatively correlated with starch metabolism.

The fatty acid synthesis module contains a pentatricopeptide repeat-containing protein (encoded by At5g50390). Because such proteins are thought to participate in modulating the stability of organelle transcripts 69, this correlation may indicate a role for this gene in coordinating nuclear-plastidic interactions for regulating fatty acid biosynthesis, via, for example, the control of the plastid-encoded subunit gene of acetyl-CoA carboxylase (accD, AtCg00500). This is the plastid-encoded biochemical function that is most directly involved in fatty acid biosynthesis.

The leucine catabolism module contains a putative transcriptional regulator, At5g49450, which may code for a bZIP family transcription factor.

These transcriptome level analyses provide insights into the hierarchical modularity of Arabidopsis metabolism. Hierarchical organization can be considered the nesting of one, tightly connected module within a larger one, still distinct and separable from the surrounding network. Hierarchical organization is particularly evident with leucine catabolism, which forms a functionally coherent and tightly connected module, nested within a supermodule [see Fig. 8 and Additional file 3].

Hierarchical organization of modules has been observed for metabolic and transcriptional networks of E. coli 707172 and metabolic networks of 42 other organisms 7374. Hierarchical modular organization has been postulated as the topology explaining the observed features of biological networks, like scale-free character, high clustering coefficient and short path lengths 7475. Such architecture may enable better adaptation of the metabolic fluxes to the changing conditions 74.

For Arabidopsis, the incorporation of leucine catabolism into an overarching catabolic module is observed both on the topological and functional level. From the topological point of view, genes participating in the leucine catabolism form a fully connected subnetwork, while they also exhibit weaker connection to the common larger set of genes. Thus, a smaller, more strongly connected module is nested within the bigger and less interconnected one. From the functional point of view, one function (leucine catabolism) forms a separable part of the larger, more encompassing biological process. The supermodule contains genes that appear to have a broader but common biological task of maintaining cellular energy balance via catabolism. Specifically, the 26 genes most highly correlated with the eight genes in the leucine catabolic module (defined in Fig. 2B) include genes that might be involved in protein turnover (At1g76410), amino acid catabolism (At1g06570, At1g08630, At2g14170, At3g47340, At5g63620, in addition to eight genes from leucine catabolism pathway), carbohydrate catabolism (At1g18270, At5g20250, At5g49360), lipid breakdown (At3g51840, At5g41080) and cell wall degradation (At4g30270). Five genes are annotated as associated with aging, senescence, or response to light or sucrose stimuli (At2g43400, At3g47340, At4g30270, At4g35770, At5g20250). Because plants are photoautotrophs, such catabolic processes may be expected to become critical when photosynthesis cannot maintain cellular energy balance, for example during senescence, seed germination, carbon deprivation or autophagy 37394276. Consistent with this concept, we find that the genes within this supermodule are co-induced to highest levels of expression under experimental conditions that are expected to limit photosynthetic carbon fixation and induce catabolic processes for maintaining energy balance (e.g., oxidative, cold, heat and drought stress – experiments from NASCArrays: NASCARRAYS-28, NASCARRAYS-79, NASCARRAYS-70; disruption of plastome-nucleome communication, NASCARRAYS-89; mutations and illumination conditions that affect the phytochrome response, NASCARRAYS-89, NASCARRAYS-124; sucrose starvation, Contento et al. 76 and NASCARRAYS-14; and hexose signaling, NASCARRAYS-40 8).

This catabolic supermodule also contains genes with well-defined or partially defined molecular functions, but undefined physiological functionality (e.g. At1g76410, zinc finger protein; At5g21170, protein kinase; At5g16340, AMP-binding protein). The inclusion of these genes within this catabolic supermodule may indicate their broader physiological function in catabolic processes; a hypothesis that can be experimentally tested by new genetic and biochemical analyses.

The experimental and meta-data was obtained from online microarray depositories: NASCArrays 43 and PLEXdb 44. The data obtained from NASCArrays, from 72 experiments comprising of 956 Affymetrix ATH1 microarray slides, had already been normalized with the Affymetrix MAS 5.0 algorithm to the common mean value of 100. We scaled the data from PLEXdb to set the mean to 100.

The reproducibility of experiments was qualitatively assessed on scatter plots by visual inspection and chips with poor biological replicates were discarded. The remaining biological replicates were averaged to yield 424 samples. Determination of biological replicates (replicates of the same treatment within the experiments) was done manually, based on the slide name and experiment information. Most slides have descriptive names that summarize treatment and the replicate status. Thus, slides that are biological replicates of the same treatment have the same slide name, differing only in "Repx" suffix, e.g.,

Shirras_1–3_Non-Calcicole_Rep1

Shirras_1–4_Non-Calcicole_Rep2

The data was subsequently normalized to the same range by a median absolute deviation (MAD)-based scale normalization method described by Yang et al., 77. Expression values x_ij on microarray chip j were multiplied by the factor CMADj MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaqcfa4aaSaaaeaacqWGdbWqaeaacqWGnbqtcqWGbbqqcqWGebardaWgaaqaaiabdQgaQbqabaaaaaaa@323F@, where MAD is defined by

MAD_j = median_i{|x_ij-median_i(x_ij)|}

and the constant C is an arithmetic mean of MAD

C = ∑ j = 1 n M A D j n MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4qamKaeyypa0tcfa4aaSaaaeaadaaeWbqaaiabd2eanjabdgeabjabdseaenaaBaaabaGaemOAaOgabeaaaeaacqWGQbGAcqGH9aqpcqaIXaqmaeaacqWGUbGBaiabggHiLdaabaGaemOBa4gaaaaa@3BDA@

The normalized data is available online 11. ATH1 probe set-to-Locus ID mapping was obtained from TAIR, March 31^st 2007 7879. The experimental information is included in Additional file 2.

The choice of genes involved in each pathway was guided by: Arabidopsis Lipid Gene Database for fatty acid synthesis genes 8081; Arabidopsis Starch Metabolism Network project 82 for starch metabolism; and original references for leucine catabolism 428384. We further considered each candidate gene based on original references.

The 126 genes used in the initial network construction (Fig. 2), including candidates for functions in fatty acid synthesis, starch metabolism and leucine catabolism, are shown in Additional file 1. A correlation matrix was calculated for these genes, using Pearson Product-Moment Coefficient, defined as

r x , y = ∑ i = 1 n ( x i − x ¯ ) ( y i − y ¯ ) ( n − 1 ) δ x δ y MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemOCai3aaSbaaSqaaiabdIha4jabcYcaSiabdMha5bqabaGccqGH9aqpjuaGdaWcaaqaamaaqahabaGaeiikaGIaemiEaG3aaSbaaeaacqWGPbqAaeqaaiabgkHiTmaanaaabaGaemiEaGhaaiabcMcaPiabcIcaOiabdMha5naaBaaabaGaemyAaKgabeaacqGHsisldaqdaaqaaiabdMha5baaaeaacqWGPbqAcqGH9aqpcqaIXaqmaeaacqWGUbGBaiabggHiLdGaeiykaKcabaGaeiikaGIaemOBa4MaeyOeI0IaeGymaeJaeiykaKIaeqiTdq2aaSbaaeaacqWG4baEaeqaaiabes7aKnaaBaaabaGaemyEaKhabeaaaaaaaa@53D2@

The values of Pearson correlations above 0.3 are statistically highly significant (p-value < 0.00001 after Bonferroni correction for multiple testing 85). We additionally assessed the validity of correlations higher than 0.6 by performing 10,000 permutations of the corresponding data vectors. Correlations that had a permutation-based p-value < 0.0001 (in our case, all pairs of genes examined) were considered for further analysis [see Additional file 4]. The networks in Fig. 2 were constructed by placing an edge between any pair of genes correlated above the threshold. Modules in the network were defined by the grouping produced by the network visualization software.

To test whether there are more correlations between the genes from the same metabolic pathway (i.e. from fatty acid biosynthesis, starch metabolism, or leucine catabolism category) than could be expected by chance, we compared the co-expression network from experimental data (Fig. 2B) to 10,000 computationally generated random networks. In these random networks the total number of the edges and the number of neighbors for each node was set to be the same as in the original graph. A function for generating random graphs was implemented in R. It constructed graphs by random filling of the adjacency matrix, where nodes' degrees were constrained and self-loops not allowed. Indeed, the proportion of within-pathway edges is significantly higher in our coexpression network compared to such proportions in random graphs (245/444 versus a mean of 112/444, which is equal to the difference of ~20 standard deviations, Fig. 3).

The calculation of Pearson correlations between the three hub-genes (At2g40840, At4g34030, At2g05990) and all other genes on the chip (Table 1) was conducted in our publicly available software, MetaOmGraph 11. MetaOmGraph is able to handle large datasets with an efficient use of memory and is integrated with other publicly available bioinformatics tools in MetNet Exchange suit 8687.

The genes belonging to catabolic supermodule [see Fig. 8 and Additional file 3] were identified by intersection of lists of all genes that are correlated above 0.5 threshold with each of eight genes from leucine catabolism module.

Three independent alleles named Atperk10-1, Atperk10-2, and Atperk10-3 corresponding to T-DNA insertions at gene AtPERK10 locus were found in the mutant collections generated at INRA-Versailles 8889, GABI-KAT 9091 and SALK 9293. The T-DNA insertions are located in 5'-UTR, exon 1 and intron 6, respectively.

Two individual plants (100–400 mg fresh weight) were boiled in 50 ml 80% (v/v) ethanol. The decolored plants were then ground in 80% ethanol with a 2 ml Konte tissue grinder, centrifuged for 10 min at 3,000 g at room temperature. The pellet was washed twice with 80% ethanol. After centrifugation, the insoluble material was suspended in 1 ml of distilled water and boiled for 30 min. Total starch was quantified using a commercial Glc assay kit (catalog no. E0207748; R-Biopharm, Darmstadt, Germany), according to instructions in the kit protocol. The assay employs amyloglucosidase to achieve complete digestion of the starch sample to Glc, and measurement of the Glc by application of a Glc oxidase reagent. The average of six biological replicates, using two individual plants for each replicate, was calculated.

The computations (normalization, correlation matrix, permutation-based p-values, random graphs generation, hypergeometric distribution) were conducted in R software 94. The RNA profiles were plotted in MetaOmGraph. The network of metabolic genes in Fig. 2 was visualized using GraphExplore 95. The pathway data is from MetNetDB database. The R code is available upon request.