JXB Advance Access originally published online on May 23, 2005
Journal of Experimental Botany 2005 56(417):1887-1896; doi:10.1093/jxb/eri179
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
RESEARCH PAPER |
Integrative gene-metabolite network with implemented causality deciphers informational fluxes of sulphur stress response
1Max Planck Institute of Molecular Plant Physiology, Department of Molecular Physiology, Am Mühlenberg 1, D-14476 Golm, Germany
2Timiryazev Institute of Plant Physiology, Russian Academy of Sciences, Botanicheskaya Str. 35, Moscow 127276, Russia
* To whom correspondence should be addressed in Germany. Fax: +49 331 567 8134. E-mail: nikiforova{at}mpimp-golm.mpg.de
Received 17 November 2004; Accepted 4 April 2005
 |
Abstract |
|---|
 Top Abstract IntroductionMaterials and methodsResults and discussionConcluding remarksReferences |
|---|
The systematic accumulation of gene expression data, although revolutionary, is insufficient in itself for an understanding of system-level physiology. In the post-genomic era, the next cognitive step is linking genes to biological processes and assembling a mosaic of data into global models of biosystem function. A dynamic network of informational flows in Arabidopsis plants perturbed by sulphur depletion is presented here. With the use of an original protocol, the first biosystem response network was reconstructed from a time series of transcript and metabolite profiles, which, on the one hand, integrates complex metabolic and transcript data and, on the other hand, possesses a causal relationship. Using the informational fluxes within this reconstruction, it was possible to link system perturbation to response endpoints. Robustness and stress tolerance, as consequences of scale-free network topology, and hubs, as potential controllers of homeostasis maintenance, were revealed. Communication paths of propagating system excitement directed to physiological endpoints, such as anthocyanin accumulation and enforced root formation were dissected from the network. An auxin regulatory circuit involved in the control of a hypo-sulphur stress response was uncovered.
Key words: Auxin, causality, metabolome, network, network topology, plait concept, scale-free network, sulphur metabolism, systems biology, transcriptome
|
Introduction |
|---|
|
TopAbstractIntroductionMaterials and methodsResults and discussionConcluding remarksReferences |
|---|
Living organisms are complex multi-elemental, multi-functional systems existing in ever-changing environments. The viability of the system is provided via flexible and effective control circuits of multiple informational fluxes interconnecting in a dense network. This hierarchically organized network of negative feedback stimuli subordinate to superior positive feedback is so fundamental that it has been proposed as a minimal but sufficient definition of life (Korzeniewski, 2001
). Unravelling such networks will allow global models of biosystem function to be characterized.
Network reconstruction and analysis are starting to be widely used to characterize and predict biosystem behaviour, giving rise to a new branch of biological knowledge, ‘network genomics’ (Forst, 2002). Until recently, such analyses have been limited to one level of manifestation of the genetic information, i.e. transcript networks (Thieffry et al., 1998, Shen-Orr et al., 2002; Rosenfeld et al., 2002; Featherstone and Broadie, 2002) or metabolic networks (Schuster et al., 2002; Fiehn and Weckwerth, 2003). However, changes in transcript levels are transferred to changes in metabolite levels and thereby to physiological endpoints via adaptations of physiology and homeostasis. Therefore, progressive system characterization involves integrating multiple levels of realization of the genetic information, for example, by superimposing transcript, protein, and metabolite profiles.
In a recent study, correlation analysis between the yeast transcriptome and proteome revealed that yeast sequesters sulphur during Cd2+ detoxification (Fauchon et al., 2002). Attempts to combine transcript and metabolic data in a correlation analysis have been undertaken by Askenazi et al. (2003), revealing pair-wise associations between fungal genes and two metabolites important for bioengineering, and by Urbanczyk-Wochniak et al. (2003), revealing genes correlating reliably to nutritionally important metabolites.
As another approach to integrate transcriptome and metabolome, data are mapped on known metabolic pathways (Grosu et al., 2002; Hirai et al., 2004). Among software tools suitable for such visualizations, AraCyc (http://www.arabidopsis.org/tools/aracyc/, Mueller et al., 2003) and MapMan (http://gabi.rzpd.de/projects/MapMan/, Thimm et al., 2004) are designed especially for Arabidopsis. The latter approach, however, is based on legacy and static metabolic data, so it generally does not allow previously unknown biochemical or regulatory pathways to be identified. The elucidation of informational fluxes linking input signals with response endpoints must be possible from a network which integrates transcriptional and metabolic changes in an unbiased way. In the course of the study presented here, an unbiased gene-metabolite network of correlations was reconstructed from transcript and metabolite profiles and the resulting informational fluxes controlling the systems response to sulphur deprivation in Arabidopsis were analysed.
|
Materials and methods |
|---|
|
TopAbstractIntroductionMaterials and methodsResults and discussionConcluding remarksReferences |
|---|
Physiological experimenting: transcript and metabolite profiling
For sulphur-starvation experiments, Arabidopsis thaliana genotype Col-0 plants were grown on a solidified agarose medium in sterile Petri dishes (half-normal Murashige-Skoog salts for control sulphur-sufficient medium, 89% less sulphur for sulphur-depleted medium). Sulphur depletion was applied as constitutive stress (germination and growth on sulphur-deficient medium for 10 d or 13 d) and induced stress (germination on normal medium for 8 d, then the transfer of seedlings to a sulphur-deficient medium for 6 d or 10 d); seedlings grown on sulphur-sufficient medium were used as the four corresponding controls. Material for each of eight experimental points was collected as whole seedlings, in five repetitive pools, containing 500–600 plants in each pool. The experiment is described in detail previously (Nikiforova et al., 2003
). The same plant material was used for RNA isolation followed by the transcript profiling technique, as well as for metabolite profiling, elemental sulphur, thiol, and anthocyanin measurements (using GC-MS, HPLS, ICP-AES, and spectrophotometry). As a result of these studies, a sulphur-deficient transcriptome (Nikiforova et al., 2003) and metabolome (Nikiforova et al., 2005) was described, providing a basis for the reconstruction of an integrated gene-metabolite network.
Network reconstruction protocol
To reconstruct a causal gene-metabolite network of statistically significant correlations, the original algorithm was elaborated (described and discussed in detail below). A condensed overview on the applied techniques is provided here. First, the transcript and metabolite profiles were combined in dataset 1 which consisted of 6454 non-redundant genes and 81 non-redundant chemical compounds, or ‘metabolites’, all containing relative concentration levels at eight experimental points. Then, the following step-by-step protocol was used to reconstruct the response gene-metabolite network from the transcript and metabolite profiles.
- From dataset 1 those genes were selected which correlated to sulphur and sulphur-responding metabolites (Fig. 1A). For this: (a) a Pearson correlation coefficient r between each of the sulphur-responding metabolites and the whole set of genes from dataset 1 was calculated. (b) The gene expression part of dataset 1 was shuffled 1000 times. (c) With the produced 1000 shuffled datasets step (i, a) was repeated 1000 times for each of the sulphur-responding metabolites. Distributions of r for the original and shuffled datasets were analysed. (d) Genes, correlating reliably to sulphur-responding metabolites, were determined from the comparisons of r distributions in the original and shuffled datasets, if they appeared in a histogram bin, that contains at least two times more genes from the original dataset than from shuffled datasets (Fig. 1A, lower graph).
- Transcript levels for a set of genes, purified from noise in step (i), were combined with a dataset of 81 metabolites into dataset 2.
- The Pearson correlation coefficient (r) distance matrix and the mutual information (MI) distance matrix of dataset 2 were calculated, transformed for compatibility (tr and tMI) and plotted (Fig. 1B, blue dots).
- Dataset 2 was shuffled.
- r distance matrix and MI distance matrix of shuffled dataset 2 were calculated, transformed for compatibility (tr and tMI), and plotted (Fig. 1B, purple dots).
- tr and tMI distance matrices of the original dataset 2 were overlaid with tr and tMI distance matrices of the shuffled dataset 2 (Fig. 1B).
- Noise correlations were cut off by setting graphically a threshold for overlaid matrices (Fig. 1B, dotted lines) in a border of the area strongly covered by associations from the shuffled dataset; as a result significant pair associations, gene/gene, gene/metabolite, or metabolite/metabolite, that were acceptable for network reconstruction were determined (Fig. 1B, shadowed area).
- A response gene-metabolite network was formed by pair-wise connection of genes or/and metabolites from all significant associations.
|
Equations and algorithms
To make calculations of Student's t-test and the Pearson correlation coefficient r, an algorithm incorporated into the Microsoft Excel 2000 software program was used.
To calculate mutual information (MI), the algorithm described by Steuer et al. (2002) was used, incorporated into the MetaGeneAlyse web tool (Daub et al., 2003), and improved with B-spline functions (Daub et al., 2004).
To threshold noise correlations, original datasets were shuffled with an over-sampling rate of 1000 with the use of an original algorithm, incorporated into the MetaGeneAlyse web tool (Daub et al., 2003).
To produce compatible correlation matrices, equations for r and MI calculations were transformed in a way to have the strongest correlations approaching 0 (tMI for transformed MI and tr for transformed r), according to the following equations, incorporated into the MetaGeneAlyse web tool was used:
 | (1) |
 | (2) |
To visualize the reconstructed network, the Pajek computer program for large network analysis (http://vlado.fmf.uni-lj.si/pub/networks/pajek/; Batagelj and Mrvar, 1998) was used.
|
Results and discussion |
|---|
|
TopAbstractIntroductionMaterials and methodsResults and discussionConcluding remarksReferences |
|---|
Integrating transcript and metabolite profiles into one dataset
For an integrated analysis, total RNA and hydrophilic metabolites were isolated from the same samples (five repetitions each; each repetition a pool of 500–600 seedlings) and used to obtain transcript profiles by array hybridization (Nikiforova et al., 2003
) and metabolite profiles by gas chromatography-mass spectrometry (GC-MS) analysis (Nikiforova et al., 2005). Relative transcript amounts for 6454 non-redundant genes and relative concentrations of 81 non-redundant chemical compounds (78 detected by GC-MS plus sulphur, glutathione, and anthocyanins measured using other methods) were determined at eight experimental points under conditions of constitutive and induced sulphur starvation, which lasted for 6–13 d. The entire compilation of these experimental data points is subsequently called dataset 1.
Reconstructing a statistically significant response network
During the developing hypo-sulphur stress, concentration levels of sulphur-responding genes and metabolites changed differentially in a time-dependent manner, forming distinct alteration patterns. Assuming that the more similar the pattern, the shorter the distance between genes and metabolites in the communication network, existing in a plant as a biosystem, the combined gene/metabolite distance matrix was calculated for dataset 1 using Pearson correlation coefficient r.
For the number of transcripts (6454) and metabolites (81) determined, the number of all possible pair-wise correlations, constituting a correlation matrix, is very large ((6454+81)2/2 equals about 21.4 million). Inherently this contains a correspondingly large noise component. To decrease dataset 1 by its putative noise component, only those genes which correlate reliably to sulphur and sulphur-related metabolites glutathione, anthocyanins, allantoin, O-acetyl-serine, putrescine, raffinose, serine, tryptophan, and uric acid (assigned as being significantly altered by sulphur depletion) were kept for further analysis. To detect these genes, the original dataset 1 was first shuffled (or randomized). From the shuffling of the dataset and recalculation of a distance matrix based on the shuffled dataset it is possible to estimate a significance threshold. The smallest distance calculated from the shuffled dataset can be regarded as the upper significance threshold for the unshuffled case. The applied shuffling algorithm of the MetaGeneAlyse tool (Daub et al., 2003) works on a dataset row by row. Within each row two values are randomly chosen and exchanged. For n values in a row an over-sampling rate of o exchanges nxo pairs of values. For dataset 1, where each row represented one gene or metabolite with corresponding eight values, an over-sampling rate of 1000 was used. The distribution of values within a row was conserved. By this procedure, 1000 shuffled datasets were generated and assumed to contain no sense information except noise. Then, correlation matrices were calculated from these 1000 shuffled datasets and compared with the correlation matrix of the original dataset (step 1 of the ‘Network reconstruction protocol’ above). A case example is shown in Fig. 1A. There, the correlation of serine with respect to all genes is shown as a function of the number of genes displaying a certain correlation coefficient. As expected, the vast majority of genes displays low r-values between –0.8 and 0.8. Distance matrices of all shuffled datasets, although not identical, produced highly similar distributions of r, which, however, were clearly different from those produced by original distance matrices, derived from the experimental data (Fig. 1A, upper graph). This observation was taken as an important argument for the robustness of the method with respect to noise.
Now, in order to filter the genes correlating reliably to sulphur and sulphur-related metabolites, the following discrimination criterion was introduced: the correlations were assumed to be reliable if the number of correlating genes in an analytical bin in a histogram of the distribution of analytical and virtual distances (Fig. 1A, lower graph) is at least 2 times higher than in a virtual bin of the shuffled dataset. Based on this assumption, those bins which were shaded in Fig. 1A were considered to contain reliably correlating items. Genes fulfilling these selection criteria were combined within a new noise-free dataset (dataset 2), which was used for response network reconstruction.
However, uneven distribution of the measured values of sulphur and some metabolites resulted in a weak relevance based on the Pearson correlation coefficient. Therefore, in addition, new distance matrices were calculated for the same dataset 2 (original and shuffled), using mutual information (MI) as another correlation measure, independent of the uneven value distribution (Butte and Kohane, 2000; Steuer et al., 2002; Daub et al., 2004). Finally, to identify significantly correlating associations acceptable for network reconstruction, it was necessary to estimate the threshold for correlation reliability by superimposing the distance matrices obtained for both approaches (i.e. Pearson correlation r and MI, for original and shuffled datasets) with the idea to consider only those associations for which both methods pointed to a reliable correlation. However, different maxima for the strongest correlations calculated with the use of r and MI prevented the direct comparison of both matrices. Therefore, to produce compatible matrices, the equations for r and MI calculations were transformed (see ‘Equations and algorithms’ above). After transformation, high correlations resulted in small values for both matrices, with the strongest correlations approaching 0 (tr for transformed r, equation 1, and tMI for transformed MI, equation 2). Now, by overlaying the calculated analytical tMI and tr matrices (Fig. 1B, blue dots) and virtual tMI and tr matrices of the shuffled dataset (Fig. 1B, purple dots), the threshold for correlation reliability was estimated graphically. The peak with tMI around 0.5, which is present in both original and shuffled datasets, contains associations with outlying values, which would be detected falsely as positive using the Pearson correlation coefficient alone. From the overlaying matrices, the area strongly covered by associations from the shuffled dataset was cut off. As a result, acceptable values for correlations to be considered in the network reconstruction were finally found as follows: tr<0.45 (|r| 
0.9) and tMI<0.3 (Fig. 1B, shaded area).
Now, based on the central assumption on item co-behaviour as introduced above (i.e. the more similar the pattern, the shorter the distance between genes and metabolites in the communication network), a response communication network was reconstructed by establishing a connecting link between paired elements, gene/gene, gene/metabolite, or metabolite/metabolite, if an association of these two elements appears among those defined as significantly correlated (shaded area in Fig. 1B). Under the chosen connectivity thresholds, the reconstructed network contained 541 elements (vertices) and 5212 associations (edges).
Global properties of the reconstructed response network suggest robustness and stress tolerance of the underlying biological system
The systems response develops by propagating through the network of informational flows reconstructed in the course of the present study. The key aspects of network functionality can be predicted directly from its structure (Albert et al., 2000; Jeong et al., 2000; Stelling et al., 2002). The network connectivities k (6.6 in average) were distributed extremely non-homogeneously. The distribution probability N(k) had a linear tail for large k, following a power law N(k) 
k–
, with the exponent =2.27 (Fig. 2A). This value lies within the range found for large networks of different nature, that is, between 2.1 and 4 (Barabasi and Albert, 1999). Such non-homogeneously-wired networks, called ‘scale-free networks’ (Barabasi and Albert, 1999), possess a number of universal characteristics. One of them is high tolerance to errors due to high robustness (Albert et al., 2000). Local errors/changes rarely lead to the loss of the global information-carrying ability of the network, as was demonstrated for both in silico and in vivo mutagenesis studies for the E. coli metabolic network (Edwards and Palsson, 2000). However, highly connected nodes (hubs) are the sites of system vulnerability in scale-free networks (McCabe, 2002; Clipsham et al., 2002). Critically important for network stability, hubs can be considered as putative controllers of homeostasis maintenance. In Table 1, vertices with the highest number of connectors (20–32) are listed. When analysing which of the different functional categories of genes shows the highest number of connectors (Table 2), it becomes clear that nucleotide metabolism, protein destination, and intracellular transport are the three most hub-rich functional categories. This finding resembles gene expression networks in E. coli (Thieffry et al., 1998), where the largest group of hubs is represented by protein synthesis and destination genes. By contrast, genes involved in protein synthesis and subclass translation displayed the lowest overall connectivity (2.5), which may result from the precise adjustment of the enzymes involved in the regulation of translation, whereas ribosomal proteins, which are able to associate with a variety of proteins, provide higher than average connectivity in the network (8.1) (Table 2).
|
|
|
Yeast's general ability to compensate for individual mutations has been proposed to be largely a result of the scale-free properties of its gene expression network (Featherstone and Broadie, 2002
). This conclusion has been extended here to a gene-metabolite communication network of a multicellular organism maintaining homeostasis during nutritional stress. The scale-free network topology suggests that the stability of the newly formed homeostasis under sulphur-limiting conditions is due to a high redundancy of gene/metabolite communication paths, triggered by critical changes in non-redundant excitement-accumulating hubs.
Implementing a causal relationship to the reconstructed response network
The characteristic feature for both applied methods of calculating correlation distance matrices (the Pearson correlation coefficient and mutual information) is non-causality: nothing in the definition of correlation implies that the relation between two variables is one of cause and effect. For the network reconstructed from these matrices, this implies that the direction of informational flows between elements is not defined (Fig. 2B). However, it is assumed here from a priori knowledge of the primary cause of system excitement, i.e. depletion of sulphate from the medium, that the changed sulphur level is the starting point, or excitation, for the response development. In keeping with this assumption, by centralizing the vertex ‘sulphur’ a general ‘cause-to-effect’ directionality of information fluxes along network paths from sulphur to distant elements has been implemented (Fig. 2C).
Mining the reconstructed network 1: combination with legacy data
In addition to its general topology and characteristics, which are indicative for properties of the given biological system, a network such as that shown in Fig. 2C contains important biological information relating to individual pathways. In this network, sulphur itself is a typical hub with a close to maximal number of direct connectors—31 (Table 1), and an even higher number of second order connectors (those which connect to direct connectors), about 200. To extract indicative routes of informational fluxes, which are directed to sulphur-responding metabolites, from such an interlaced net certain branches of connectivity were selected with the following procedure. In the upper region with regard to a certain metabolite (i.e. between sulphur and the metabolite) only those links belonging to a path ‘sulphur—the metabolite’ were left; in the lower region all existing links were left (Fig. 3A).
|
Applying the same unravelling procedure, another, now hormone-related fragment was isolated from the network (Fig. 3B).
Mining the extracted fragments provided examples of where the reconstructed network confirms and extends existing data, or leads to new knowledge. For a case in point, the positions of O-acetyl-serine (OAS), an important indicator of sulphate deprivation in plants (Nikiforova et al., 2003; Hirai et al., 2003), and N-acetyl-serine (NAS) were investigated. NAS is derived from OAS by non-enzymatic intramolecular O- to N-acyl migration (Flavin and Slaughter, 1965). Their relative concentrations are highly interdependent, and so both metabolites have to show coupled alterations in relative concentration levels, as is indeed reflected in their close network positioning: OAS and NAS appear in the network as closely connected vertices, equally distant from sulphur (connectors of the 2nd order, marked with green, Fig. 3A).
As another illustration, the flows of information processing regarding tryptophan were examined. In a previous study, it was suggested that the increase in tryptophan observed in sulphur-deprived plants could be explained by assuming that the reduced level of cysteine leads to a surplus in serine, which is subsequently converted into tryptophan (Nikiforova et al., 2003). It has to be emphasized that this hypothesis was based on legacy data regarding biochemical pathways (for an overview of amino acid biosynthesis in plants, see Coruzzi and Last, 2000). In this respect, it is remarkable that the reconstructed network which was created solely on ungoverned analysis of co-behaviour of genes and metabolites displays only one path from sulphur to tryptophan, and this path runs via serine (Fig. 3A), therefore corroborating the hypothesis raised previously. Further, it is worth mentioning that the causal directionality ‘sulphur–serine–tryptophan’ could only be detected due to the centralized position of sulphur, which was implemented to the reconstructed network as a priori knowledge of the system exciter. This example was regarded as a striking case for the usefulness of the reconstructed network as a concept tool for identifying biological paths of communication.
Mining the reconstructed network 2: finding paths from system exciter to physiological endpoints
The systemic response of plants to sulphur deprivation results in several endpoints of biochemical pathways and physiological reactions, such as increased root formation and accumulation of anthocyanins (Nikiforova et al., 2003). Alhough both physiological events are typical reactions to stress (Malamy and Ryan, 2001; Steyn et al., 2002; Lopez-Bucio et al., 2003), both do develop in sulphur-starved plants and thus they were traced in the reconstructed network.
As shown in Fig. 3A, vertex ‘anthocyanins’ displays only two links, and both are upstream-positioned. Among all the metabolites, this was the only one that did not possess any link to parallel or downstream vertices. Such positioning is considered as a path leading from the system exciter to a communication endpoint, in agreement with the identification of anthocyanins as a physiological endpoint of stress response.
As an example of new biological information resulting from the reconstructed network analysis, a path to another physiological endpoint is presented, i.e. enhanced root formation, represented by an auxin spindle (Fig. 3B). Starting from sulphur, several interlacing redundant paths pass nitrilase 3 and a putative myrosinase-associated protein, both involved in auxin biosynthesis, and then lead via the auxin signal transduction factor calmodulin 3 to the hub-forming node IAA28, an auxin-related transcriptional factor, known to be mainly expressed in roots (Rogg et al., 2001). Analysis of IAA28 expression history in the Stanford Microarray Database (http://www.stanford.edu/microarray; Gollub et al., 2003) provided experimental evidence that IAA28 is, first, involved in the auxin signal transduction cascade, and second, in a conditional manner: the IAA28 gene was significantly over-expressed after exogenous auxin treatment only in the background of the Arabidopsis IAA24 mutant (described in Przemeck et al., 1996; Hardtke and Berleth, 1998; Berleth et al., 2000) (experiment IDs 26716-26717, experimenter T Berleth). It should be stressed that this distinction has not been articulated before, even though the array data have long been publicly available. With this example, it was demonstrated that a hypothesis created by reconstructed networks can seed specific searches for verification in array databases. In this case, the data specifically derived from the IAA24 mutant profiles are in full agreement with the supposed role of IAA28 in auxin signalling. It is important to emphasize that, with the increasing growth of publicly available databases, such experimental verification of network-derived hypotheses, arising from, for example, mutants, will become of increasing importance.
The detected auxin spindle extends previous observations regarding auxin-related signalling (Fig. 4). Sulphur-deficient homeostasis causes a surplus metabolic flux via auxin and the activation of auxin-induced genes (Hirai et al., 2003; Nikiforova et al., 2003). The altered auxin content triggers changes in free calcium levels in plant cells (Felle, 1988; Gehring et al., 1990; Kalra and Bhatla, 1999), which are sensed by calmodulin. The interaction between auxin and calmodulin has been confirmed experimentally (Naren et al., 1995; Okamoto et al., 1995; Choi et al., 1996; Yang and Poovaiah, 2000), thus proving the relevance of this part of the path derived from the network. As a novel finding extending this path it is proposed that activated calmodulin influences the expression of the IAA28 gene, based on their relative positions within the reconstructed sulphur response network. In turn, IAA28 represses the transcription of auxin-induced genes, as shown by a gain-of-function mutation in IAA28 (Rogg et al., 2001). Thus, by incorporating the network-derived interaction between calmodulin and IAA28 the auxin signalling path has been extended to a closed circuit (Fig. 4), thus providing a new example of the feedback control of multiple informational fluxes, and new evidence for the fundamental principle of a living system as a network of inferior negative feedback subordinate to a superior positive feedback.
|
At first sight, transcriptional repression of auxin-induced genes under conditions of depleted nutrient supply, which fosters exploitation of new soil layers by enhanced root growth, seems difficult to reconcile. However, the metabolic overflow to auxin caused by sulphur deficiency via the path ‘sulphur–serine–tryptophan’ (proposed by Nikiforova et al., 2003
, and detected as an information flux in the reconstructed hypo-sulphur response network, present study) may well require negative feedback control over the expression of auxin-induced genes, which can be provided via the activity of IAA28. For the overall survival strategy of a plant as a system, feedback auxin regulation can serve as a way to save resources under conditions in which further root enforcement leads to fatal resource expenditure and, as a result, the inability to set seeds for a new generation.
Mining the reconstructed network 3: check on reliability
The implemented causal directionality allows hub input (upstream positioned) and output (downstream positioned) signals to be distinguished. If a hub is formed by a gene involved in transcriptional regulation, then in comparative promoter analysis hub output genes are expected to contain common motifs in their promoter regions. This expectation comes solely from the network topology analysis and therefore can serve as a check for reliability in network reconstruction. As an example, statistical motif analysis was undertaken in promoter sequences of IAA28 input and output genes, as hub-forming IAA28 possesses the highest number of the direct connectors among the genes with annotated transcription factor activities (Table 1). Using the tool for motif analysis available at the TAIR web page (http://www.arabidopsis.org/tools/bulk/motiffinder/index.jsp), twice as many common motifs were found in the queried output, than input, promoters. This result fits the expectation from our network topology, and thus supports causal directionality of the reconstructed network.
Hub-containing networks support the plait concept
The reconstructed response network also highlights the distinctions between specific and common parts of response routes. According to the plait concept first proposed by Nikiforova et al. (2003), the transduction of specificity during stress responses within a non-specific signalling stream is provided by the specific interlacing of biosynthetic pathways. In the example of the auxin-related spindle (Fig. 3B), one can follow the specific interlacing of correlation paths, some of which pass calcium-sensing calmodulin. Calcium-related signalling systems are well known to be redundant (Bowler and Fluhr, 2000; Yang and Poovaiah, 2002), and involvement of calcium as an intermediate signalling molecule in responses to diverse stresses necessitates its interactions in a combinatorial fashion, resulting in a dense network of interactions. Thus, calcium among other intermediate signalling molecules can be considered as an element of the main non-specific signalling stream. Such an intermediate positioning in a signal transduction cascade may lead to cross-tolerance against multiple stresses (Bowler and Fluhr, 2000). However, a specific signal should first reach the general non-specific stream. As strands of hair interlace into a plait, biological specificity at the beginning of the signalling path connects with multiple (environmental) signals at the end, thus providing the most effective, specific reaction. End responses are connected by a non-specific main signalling stream, which contains common signalling molecules, like calcium or activated oxygen species (discussed by Bowler and Fluhr, 2000), and which probably provides multiple crosstalk connections between pathways, as described in plants (Genoud and Metraux, 1999). The analysis of the reconstructed response network showed that, besides biosynthetic pathways, other correlation paths of either known, or of as yet unknown, natures provide a flow of information from initial excitement to physiological endpoints. As an extension of the plait concept, those sulphur-specific elements belonging to the upper non-interlaced part of the plait can now be dissected. In the network, these are probably the vertices between sulphur and calmodulin, and among them nitrilase 3 was indeed shown to be highly specific for sulphur stress response (Kutz et al., 2002; Hirai et al., 2003; Nikiforova et al., 2003).
|
Concluding remarks |
|---|
|
TopAbstractIntroductionMaterials and methodsResults and discussionConcluding remarksReferences |
|---|
The described approach for reconstructing gene-metabolite communication networks with implemented causality opens many research opportunities on systems behaviour. Indispensability of the hubs for information processing suggests examining alternative networks in which hub genes are knocked out. In the search for response specificity, the comparison of response networks of, for example, other nutritional stresses, will allow the identification of nutrient-specific response elements.
High-precision molecular tracking of resolved informational fluxes provides potential gains in the understanding of many responses, including stress, disease, and therapeutic effects. The advantages of this innovative approach can be demonstrated as applied to stress biology or biomedicine. When a stress/disease agent is known, tracing of signalling pathways to response endpoints through the reconstructed network as a whole, and the identification of control hubs through the analysis of network topology will highlight potential control points and lead to target drug discovery. If the stress/disease agent is unknown, causality can be implemented from downstream effects (endpoint reactions) and may lead to the identification of the agent. Taken together, reconstruction of the gene/metabolite network with implemented causal directionality provides an extension towards a consistent development of systems biology.
|
Acknowledgements |
|---|
We thank Megan McKenzie for manuscript proofreading. This work was supported by the EU commission through funding of FP5 project QLRT-2000-00103 and by the Max Planck Society.
|
Footnotes |
|---|
Present address: Center for Genomics and Bioinformatics, Karolinska Institutet, S-171 77 Stockholm, Sweden. 
|
References |
|---|
|
TopAbstractIntroductionMaterials and methodsResults and discussionConcluding remarksReferences |
|---|
Albert R, Jeong H, Barabasi AL. 2000. Error and attack tolerance of complex networks. Nature 406, 378–382.[CrossRef][Medline]
Askenazi M, Driggers EM, Holtzman DA, et al. 2003. Integrating transcriptional and metabolite profiles to direct the engineering of lovastatin-producing fungal strains. Nature Biotechnology 21, 150–156.[CrossRef][Web of Science][Medline]
Barabasi AL, Albert R. 1999. Emergence of scaling in random networks. Science 286, 509–512.
Batagelj V, Mrvar A. 1998. Pajek—Program for Large Network Analysis. Connections 21, 47–57.
Berleth T, Mattsson J, Hardtke CS. 2000. Vascular continuity, cell axialization and auxin. Plant Growth Regulation 32, 173–185.[CrossRef]
Bowler C, Fluhr R. 2000. The role of calcium and activated oxygens as signals for controlling cross-tolerance. Trends in Plant Science 5, 241–246.[CrossRef][Web of Science][Medline]
Butte AJ, Kohane IS. 2000. Mutual information relevance networks: functional genomic clustering using pairwise entropy measurements. Pacific Symposium on Biocomputing 5, 415–426.
Choi YJ, Cho EK, Lee SI, Lim CO, Gal SW, Cho MJ, An GH. 1996. Developmentally regulated expression of the rice calmodulin promoter in transgenic tobacco plants. Molecular Cell 6, 541–546.
Clipsham R, Zhang YH, Huang BL, McCabe ERB. 2002. Genetic network identification by high density, multiplexed reversed transcriptional (HD-MRT) analysis in steroidogenic axis model cell lines. Molecular Genetics and Metabolism 77, 159–178.[CrossRef][Web of Science][Medline]
Coruzzi G, Last R. 2000. Amino acids. In: Buchanan BB, Gruissem W, Jones RL, eds. Biochemistry and molecular biology of plants. Rockville, Maryland: American Society of Plant Physiologists, 358–410.
Daub CO, Kloska S, Selbig J. 2003. MetaGeneAlyse: analysis of integrated transcriptional and metabolite data. Bioinformatics 19, 2332–2333.
Daub CO, Steuer R, Selbig J, Kloska S. 2004. Estimating mutual information using B-spline functions: an improved similarity measure for analysing gene expression data. BMC Bioinformatics 5, 118.[CrossRef][Medline]
Edwards JS, Palsson BO. 2000. The Escherichia coli MG1655 in silico metabolic genotype: its definition, characteristics, and capabilities. Proceedings of the National Academy of Sciences, USA 97, 5528–5533.
Fauchon M, Lagniel G, Aude JC, Lombardia L, Soularue P, Petat C, Marguerie G, Sentenac A, Werner M, Labarre J. 2002. Sulfur sparing in the yeast proteome in response to sulfur demand. Molecular Cell 9, 713–723.[CrossRef][Web of Science][Medline]
Featherstone DE, Broadie K. 2002. Wrestling with pleiotropy: genomic and topological analysis of the yeast gene expression network. Bioessays 24, 267–274.[CrossRef][Web of Science][Medline]
Felle H. 1988. Auxin causes oscillations of cytosolic free calcium and pH in Zea mays coleoptiles. Planta 174, 495–499.[CrossRef]
Fiehn O, Weckwerth W. 2003. Deciphering metabolic networks. European Journal of Biochemistry 270, 579–588.[Web of Science][Medline]
Flavin M, Slaughter C. 1965. Synthesis of the succinic ester of homoserine, a new intermediate in the bacterial biosynthesis of methionine. Biochemistry 4, 1370–1375.[CrossRef][Medline]
Forst CV. 2002. Network genomics: a novel approach for the analysis of biological systems in the post-genomic era. Molecular Biology Reporter 29, 265–280.[CrossRef]
Gehring CA, Irving HR, Parish RW. 1990. Effects of auxin and abscisic acid on cytosolic calcium and pH in plant cells. Proceedings of the National Academy of Sciences, USA 87, 9645–9649.
Genoud T, Metraux JP. 1999. Crosstalk in plant cell signalling: structure and function of the genetic network. Trends in Plant Science 4, 503–507.[CrossRef][Web of Science][Medline]
Gollub J, Ball CA, Binkley G, et al. 2003. The Stanford microarray database: data access and quality assessment tools. Nucleic Acids Research 31, 94–96.
Grosu P, Townsend JP, Hartl DL, Cavalieri D. 2002. Pathway processor: a tool for integrating whole-genome expression results into metabolic networks. Genome Research 12, 1121–1126.
Hardtke CS, Berleth T. 1998. The Arabidopsis gene MONOPTEROS encodes a transcription factor mediating embryo axis formation and vascular development. EMBO Journal 17, 1405–1411.[CrossRef][Web of Science][Medline]
Hirai MY, Fujiwara T, Awazuhara M, Kimura T, Noji M, Saito K. 2003. Global expression profiling of sulfur-starved Arabidopsis by DNA macroarray reveals the role of O-acetyl-L-serine as a general regulator of gene expression in response to sulfur nutrition. The Plant Journal 33, 651–663.[CrossRef][Web of Science][Medline]
Hirai MY, Yano M, Goodenowe DB, Kanaya S, Kimura T, Awazuhara M, Arita M, Fujiwara T, Saito K. 2004. Integration of transcriptomics and metabolomics for understanding of global responses to nutritional stresses in Arabidopsis thaliana. Proceedings of the National Academy of Sciences, USA 101, 10205–10210.
Jeong H, Tombor B, Albert R, Oltval ZN, Barabasi AL. 2000. The large-scale organization of metabolic networks. Nature 407, 651–654.[CrossRef][Medline]
Kalra G, Bhatla SC. 1999. Distribution of membrane-bound calcium and activated calmodulin in cultured protoplasts of sunflower (Helianthus annuus L.). Current Science 76, 1580–1584.
Korzeniewski B. 2001. Cybernetic formulation of the definition of life. Journal of Theoretical Biology 209, 275–286.[CrossRef][Web of Science][Medline]
Kutz A, Müller A, Hennig P, Kaiser WM, Piotrowski M, Weiler EW. 2002. A role for nitrilase 3 in the regulation of root morphology in sulphur-starving Arabidopsis thaliana. The Plant Journal 30, 95–106.[CrossRef][Web of Science][Medline]
Lopez-Bucio J, Cruz-Ramirez A, Herrera-Estrella L. 2003. The role of nutrient availability in regulating root architecture. Current Opinion in Plant Biology 6, 280–287.[CrossRef][Web of Science][Medline]
Malamy JE, Ryan KS. 2001. Environmental regulation of lateral root initiation in Arabidopsis. Plant Physiology 127, 899–909.
McCabe ERB. 2002. Vulnerability within a robust complex system-DAX-1 mutations and steroidogenic axis development. Journal of Clinical Endocrinology and Metabolism 87, 41–43.
Mueller LA, Zhang PF, Rhee SY. 2003. AraCyc: a biochemical pathway database for Arabidopsis. Plant Physiology 132, 453–460.
Naren A, Prasad TG, Sashidhar VR, Kumar MU. 1995. Involvement of calcium in brassinolide and auxin-induced cell elongation. Current Science 69, 777–780.
Nikiforova V, Freitag J, Kempa S, Adamik M, Hesse H, Hoefgen R. 2003. Transcriptome analysis of sulfur depletion in Arabidopsis thaliana: interlacing of biosynthetic pathways provides response specificity. The Plant Journal 33, 633–650.[CrossRef][Web of Science][Medline]
Nikiforova VJ, Kopka J, Tolstikov V, Fiehn O, Hopkins L, Hawkesford MJ, Hesse H, Hoefgen R. 2005. Systems rebalancing of metabolism in response to sulfur deprivation, as revealed by metabolome analysis of Arabidopsis plants. Plant Physiology 138, 304–318.
Okamoto H, Tanaka Y, Sakai S. 1995. Molecular cloning and analysis of the cDNA for an auxin-regulated calmodulin gene. Plant Cell Physiology 36, 1531–1539.
Przemeck GKH, Mattsson J, Hardtke CS, Sung ZR, Berleth T. 1996. Studies on the role of the Arabidopsis gene MONOPTEROS in vascular development and plant cell axialization. Planta 200, 229–237.[Web of Science][Medline]
Rogg LE, Lasswell J, Bartel B. 2001. A gain-of-function mutation in IAA28 suppresses lateral root development. The Plant Cell 13, 465–480.
Rosenfeld N, Elowitz MB, Alon U. 2002. Negative autoregulation speeds the response times of transcription networks. Journal of Molecular Biology 323, 785–793.[CrossRef][Web of Science][Medline]
Schuster S, Klamt S, Weckwerth W, Moldenhauer F, Pfeiffer T. 2002. Use of network analysis of metabolic systems in bioengineering. Bioprocess and Biosystems Engineering 24, 363–372.[CrossRef]
Shen-Orr SS, Milo R, Mangan S, Alon U. 2002. Network motifs in the transcriptional regulation network of Escherichia coli. Nature Genetics 31, 64–68.[CrossRef][Web of Science][Medline]
Stelling J, Klamt S, Bettenbrock K, Schuster S, Gilles ED. 2002. Metabolic network structure determines key aspects of functionality and regulation. Nature 420, 190–193.[CrossRef][Medline]
Steuer R, Kurths J, Daub CO, Weise J, Selbig J. 2002. The mutual information: detecting and evaluating dependencies between variables. Bioinformatics 18, S231–S240.[Abstract]
Steyn WJ, Wand SJE, Holcroft DM, Jacobs G. 2002. Anthocyanins in vegetative tissues: a proposed unified function in photoprotection. New Phytologist 155, 349–361.[CrossRef]
Thieffry D, Huerta AM, Perezrueda E, Colladovides J. 1998. From specific gene regulation to genomic networks: a global analysis of transcriptional regulation in Eschherichia coli. Bioessays 20, 433–440.[CrossRef][Web of Science][Medline]
Thimm O, Bläsing O, Gibon Y, Nagel A, Meyer S, Krüger P, Selbig J, Müller LA, Rhee SY, Stitt M. 2004. MapMan: a user-driven tool to display genomics data sets onto digrams of metabolic pathways and other biological processes. The Plant Journal 37, 914–939.[CrossRef][Web of Science][Medline]
Urbanczyk-Wochniak E, Luedemann A, Kopka J, Selbig J, Roessner-Tunali U, Willmitzer L, Fernie AR. 2003. Parallel analysis of transcript and metabolic profiles: a new approach in systems biology. EMBO Reports 4, 989–993.[CrossRef][Web of Science][Medline]
Yang TB, Poovaiah BW. 2000. Molecular and biochemical evidence for the involvement of calcium/calmodulin in auxin action. Journal of Biological Chemistry 275, 3137–3143.
Yang TB, Poovaiah BW. 2002. A calmodulin-binding/CGCG box DNA-binding protein family involved in multiple signalling pathways in plants. Journal of Biological Chemistry 277, 45049–45058.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  F. Mounet, A. Moing, V. Garcia, J. Petit, M. Maucourt, C. Deborde, S. Bernillon, G. Le Gall, I. Colquhoun, M. Defernez, et al. Gene and Metabolite Regulatory Network Analysis of Early Developing Fruit Tissues Highlights New Candidate Genes for the Control of Tomato Fruit Composition and Development Plant Physiology, March 1, 2009; 149(3): 1505 - 1528. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
A. Fait, K. Hanhineva, R. Beleggia, N. Dai, I. Rogachev, V. J. Nikiforova, A. R. Fernie, and A. Aharoni Reconfiguration of the Achene and Receptacle Metabolic Networks during Strawberry Fruit Development Plant Physiology, October 1, 2008; 148(2): 730 - 750. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 Y. Noguchi, N. Shikata, Y. Furuhata, T. Kimura, and M. Takahashi Characterization of dietary protein-dependent amino acid metabolism by linking free amino acids with transcriptional profiles through analysis of correlation Physiol Genomics, August 1, 2008; 34(3): 315 - 326. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 S. Ma, Q. Gong, and H. J. Bohnert An Arabidopsis gene network based on the graphical Gaussian model Genome Res., November 1, 2007; 17(11): 1614 - 1625. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
N. Yoshimoto, E. Inoue, A. Watanabe-Takahashi, K. Saito, and H. Takahashi Posttranscriptional Regulation of High-Affinity Sulfate Transporters in Arabidopsis by Sulfur Nutrition Plant Physiology, October 1, 2007; 145(2): 378 - 388. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 K. Aoki, Y. Ogata, and D. Shibata Approaches for Extracting Practical Information from Gene Co-expression Networks in Plant Biology Plant Cell Physiol., March 1, 2007; 48(3): 381 - 390. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
F. Carrari, C. Baxter, B. Usadel, E. Urbanczyk-Wochniak, M.-I. Zanor, A. Nunes-Nesi, V. Nikiforova, D. Centero, A. Ratzka, M. Pauly, et al. Integrated Analysis of Metabolite and Transcript Levels Reveals the Metabolic Shifts That Underlie Tomato Fruit Development and Highlight Regulatory Aspects of Metabolic Network Behavior Plant Physiology, December 1, 2006; 142(4): 1380 - 1396. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||