Snapshot of iron response in Shewanella oneidensis by gene network reconstruction

Background Iron homeostasis of Shewanella oneidensis, a γ-proteobacterium possessing high iron content, is regulated by a global transcription factor Fur. However, knowledge is incomplete about other biological pathways that respond to changes in iron concentration, as well as details of the responses. In this work, we integrate physiological, transcriptomics and genetic approaches to delineate the iron response of S. oneidensis. Results We show that the iron response in S. oneidensis is a rapid process. Temporal gene expression profiles were examined for iron depletion and repletion, and a gene co-expression network was reconstructed. Modules of iron acquisition systems, anaerobic energy metabolism and protein degradation were the most noteworthy in the gene network. Bioinformatics analyses suggested that genes in each of the modules might be regulated by DNA-binding proteins Fur, CRP and RpoH, respectively. Closer inspection of these modules revealed a transcriptional regulator (SO2426) involved in iron acquisition and ten transcriptional factors involved in anaerobic energy metabolism. Selected genes in the network were analyzed by genetic studies. Disruption of genes encoding a putative alcaligin biosynthesis protein (SO3032) and a gene previously implicated in protein degradation (SO2017) led to severe growth deficiency under iron depletion conditions. Disruption of a novel transcriptional factor (SO1415) caused deficiency in both anaerobic iron reduction and growth with thiosulfate or TMAO as an electronic acceptor, suggesting that SO1415 is required for specific branches of anaerobic energy metabolism pathways. Conclusion Using a reconstructed gene network, we identified major biological pathways that were differentially expressed during iron depletion and repletion. Genetic studies not only demonstrated the importance of iron acquisition and protein degradation for iron depletion, but also characterized a novel transcriptional factor (SO1415) with a role in anaerobic energy metabolism.


Background
Iron is an important nutrient for bacteria, serving as a cofactor for proteins involved in respiration, the tricarboxylic acid (TCA) cycle, enzyme catalysis, gene regulation, photosynthesis, N 2 fixation, methanogenesis, H 2 production and consumption, oxygen transport, and DNA biosynthesis [1][2][3]. Iron exists in two redox states under physiological conditions: the insoluble Fe(III) ferric form and the relatively soluble Fe(II) ferrous form, which renders it a useful enzyme prosthetic group [4]. However, iron can induce oxidative stress by catalyzing Fenton reactions, and the prevalent Fe(III) form has low bioavailability [4,5]. Intracellular levels of iron must, therefore, be carefully controlled to meet the metabolic needs of the cell while limiting cellular damage due to iron overload [6][7][8].
Microbial iron response has been well-studied in Escherichia coli, which adapts to low iron conditions by inducing iron-binding transporters embedded in the outer and inner membranes to import iron [1,9]. Ironchelating siderophores also are secreted to solubilize iron prior to transportation across the membranes [4]. Concurrently, production of non-essential proteins that use iron is inhibited, which increases the pool of free iron in the cell [10]. Iron homeostasis in a diverse group of prokaryotes is maintained by the global transcriptional factor Fur (Ferric Uptake Regulator) and small regulatory RNA RyhB. Fur is an iron-responsive protein dimer with an aminoterminal helix-turn-helix DNA binding domain and a metal-binding domain [1,11]. Each Fur monomer complexes with Fe(II) and binds the major groove of iron responsive gene promoters at a conserved 19-bp inverted repeat sequence (GATAATGATAATCATTATC) called the "Fur box". The Fur box, located near the cognate promoter [11,12], effectively prevents recruitment of RNA polymerase holoenzyme to the promoter and thus represses transcription [13,14]. Fur directly represses the expression of RyhB small RNA, which in turn represses a large group of mRNAs encoding non-essential, iron-using proteins [10,15]. This regulatory cascade of Fur and RyhB is believed to ensure that iron is directed towards essential iron-using proteins during iron limitation [16].
Studying iron homeostasis in S. oneidensis is of interest for a number of reasons. Known terminal electron acceptors of S. oneidensis include a variety of inorganic and organic compounds such as fumarate, nitrate, thiosulfate, trimethylamine N-oxide (TMAO), Fe(III), Mn(IV), Cr(VI), and U(VI) [17]. The ability of S. oneidensis to respire Fe(III) is unique among the known γ-proteobacteria. Iron not only acts as a cofactor in S. oneidensis, but also as an important terminal electron acceptor. S. oneidensis is also striking for its high cellular demand for iron due to a high cellular content of heme [18], which is a protein cofactor requiring iron. The characteristic pink or red color of the cells indicates the high content of heme in S. oneidensis. In addition, S. oneidensis has been shown to be a fish pathogen as well as an opportunistic human pathogen [19,20]. Iron-regulated metabolisms are essential for the virulence of Vibrio cholerae [21], Bacillus cereus [22], Neisseria meningitidis [23] and Shigella species [24]. It is possible that iron may play a similar role in S. oneidensis as it does in other pathogenic γ-proteobacteria (e.g. Vibrio and Shigella).
An ortholog of Fur has been identified in S. oneidensis [2]. Its major role is to act as a repressor that mediates transcription of genes involved in siderophore biosynthesis and iron acquisition systems [2,3,25] and hence plays a role in response to iron depletion. It remains unclear, however, which other biological pathways respond to the change of iron concentration, as well as the nature of those responses. To identify iron-responsive processes, we carried out an integrative study with physiological, transcriptomic and genetic approaches.

Physiological response of MR-1 to changes in iron concentration
To examine the iron response of S. oneidensis, the wildtype strain MR-1 was grown in LB medium with different concentrations of the iron chelator 2,2'-dipyridyl used to deplete iron. LB medium, a rich source of iron (~17 μM) [26], was used to provide sufficient biomass for the following microarray experiments. Iron depletion imposed a challenge to cell proliferation and survival, resulting in an extended lag phase, slower growth rate at mid-log phase, and lower cell density at stationary phase (Fig. 1A). While 80 μM 2,2'-dipyridyl had a marginal effect on the growth rate of MR-1, the cells displayed clear growth inhibition with 160 μM and 240 μM 2,2'-dipyridyl. When 320 μM 2,2'-dipyridyl was used, cell growth was completely arrested throughout five days of observation.
To test whether iron depletion led to a rapid physiological change, MR-1 was grown to mid-log phase (OD 600 = 0.6), then 2,2'-dipyridyl was added to a final concentration of 160 μM, followed by addition of ferrous sulfate after one hour. Iron depletion rapidly slowed growth within an hour of adding 2,2'-dipyridyl. Growth was recovered by the subsequent addition of ferrous sulfate (Fig. 1B).

Transcriptomics and gene co-expression network
For microarray experiments, MR-1 cultures at mid-log phase were treated with 2,2'-dipyridyl. After one hour, iron was repleted with the addition of ferrous sulfate to the medium. 160 μM 2,2'-dipyridyl was used for iron depletion for the microarray experiments because it inhibited, but did not abrogate, cell growth (Fig. 1). Samples were collected at multiple time points during iron deple-(A) Growth curves of S. oneidensis MR-1 in liquid LB medium with different concentrations of iron chelator (2,2'-dipyridyl) Figure 1 (A) Growth curves of S. oneidensis MR-1 in liquid LB medium with different concentrations of iron chelator (2,2'-dipyridyl). Triplicates of cells grown to mid-log phase were diluted 1:100 in fresh LB liquid medium, and cell density was measured every half hour for five consecutive days. The standard deviations are shown as error bars. (B) When MR-1 cultures were grown to mid-log phase (O.D. 600 nm = 0.6), 160 μM 2,2'-dipyridyl was added to inhibit bacterial growth, which was subsequently reversed by repleting iron with 200 μM FeSO 4 . The arrows represents the time when 2,2'-dipyridyl or iron was added. Data are averages for triplicate cultures.
tion and repletion and were used for global transcriptomic analyses. The reliability of the microarray data was validated by quantitative RT-PCR of selected genes (Table  1). A high correlation coefficient of 0.98 was observed between RT-PCR and the microarray results.
The number of genes showing significant differential gene expression at each time point is listed in Table 2. The identification of 21 genes induced within one minute of iron depletion, many of which encode iron acquisition systems, clearly indicates that iron response is a very rapid process in S. oneidensis. Also notable is the fact that the total numbers of differentially expressed genes are smaller at earlier time points of iron depletion and repletion than those at later time points, at which other effects (such as accumulation of intermediate products and change in growth rate) may impact gene expression profiles. As is common in microarray datasets, a large portion of the upor down-regulated genes corresponds to genes with unknown function, indicating a much broader iron stimulon and regulon than have been deduced solely on the basis of gene annotation. When the genes with significant changes were classified according to their functional categories, several groups were predominantly represented: transport and binding proteins (mostly iron acquisition systems), regulatory functions, energy metabolism, adap-tation to atypical conditions, and biosynthesis of cofactors or proteins, as exemplified at 40 minutes of iron depletion and repletion (Fig. 2).
To gain more insights into the iron-responsive genes, a gene co-expression network containing a total of 128 genes partitioned into twelve modules (Additional file 1) was constructed (see Methods for details). Further analyses indicated that each module contains a functionally coherent set of genes. Large modules of genes related to iron acquisition, anaerobic energy metabolism and protein degradation are the most predominant in the network, suggesting that they are the major iron-responsive biological pathways in S. oneidensis. In addition, gene modules of DNA metabolism, sulfur metabolism, protein synthesis and aerobic energy metabolism also were identified.

Module of iron acquisition systems
The large module shown in Fig. 3A is composed almost exclusively of iron acquisition systems, suggesting that these genes are highly co-regulated under iron depletion and repletion conditions. These genes include the TonB system (TonB1-ExbB1-ExbD1), the hemin ABC transporter (HmuT, HmuU and HmuV), the siderophore synthetic protein (AlcA), and an iron-regulated outer membrane protein (IrgA). A majority of genes were strongly induced by iron depletion and repressed by iron repletion (see Table 3 for details). Notably, two homologous TonB systems located at the inner membrane have been annotated in S. oneidensis. The TonB1-exbB1-exbD1 operon was induced 2.4-64.6 fold during iron depletion and then repressed up to 100 fold by iron repletion, while TonB2-exbB2-exbD2 was not regulated by iron. Consistently, a "Fur box" was identified upstream of the TonB1-exbB1-exbD1 operon, but not upstream of the TonB2-exbB2-exbD2 operon [3]. In contrast, three genes (Bfr2, SO3065 and SuhB) display expression patterns opposite to the iron acquisition systems (Fig. 3A). The presence of Bfr2 can be explained as a bacterioferritin subunit that is regulated in the opposite way to iron acquisition.
Five genes (SO0448, SO0449, SO1188, SO2736 and SO3407) encoding unknown conserved inner or outer membrane proteins are regulated by iron, raising the possibility of novel iron acquisition systems. In addition, a single transcriptional regulator (SO2426) clusters with iron acquisition systems. Very recently, an independent genetic study showed that SO2426 regulated siderophore production, confirming its role in regulating iron acquisition [27].
To identify the possible transcriptional regulator for genes in this module, their upstream intergenic regions were searched for common motifs using the Gibbs Recursive Sampler [28]. A motif almost identical to the predicted "Fur box" in S. oneidensis [3] was identified (Fig. 3A). Thus, the genes in this module appeared to be directly controlled by Fur. Consistently, most of these genes were de-repressed in a fur deletion mutant [3,25].

Module of anaerobic energy metabolism
In general, genes in this module are repressed by iron depletion and induced by repletion (see Table 3 for details). This module contains many genes contributing to anaerobic energy metabolism, including c-type cytochromes, which function in anaerobic metal reduction (e.g., CymA, MtrA, MtrC and OmcA); anaerobic ribonucleoside-triphosphate reductase (NrdD); the FAD-binding subunit of the inner membrane respiratory complex (FrdA); alcohol dehydrogenase (AdhE); biosynthetic genes of cytochromes such as the heme exporter (HemE and HemH-1); and cytochrome maturation systems (CcmA, CcmB and CcmC) (Fig. 3B). In addition, ten transcriptional factors are clustered with the functional genes, implicating that they function to regulate this process.

Module of protein degradation
A module of protein degradation was identified in the network (Fig. 3C). These genes generally are induced by iron depletion and repressed by iron repletion (Table 3). They include heat shock proteins (DnaK, DnaJ and HtpG) and heat shock protein 70 (Hsp70) system-interacting proteins (IscU and IscA) [32]. An RpoH (σ 32 ) binding site (CTTGAAA and CCCCAT) was identified -35 and -10 upstream of genes in this module, which was consistent with previous observations in E. coli that RpoH was responsible for regulation of heat shock proteins [33]

Mutant analyses
To test the importance of iron acquisition genes under iron depleted conditions, a putative alcaligin biosynthesis protein SO3032 was selected for examination. A mutant of SO3032 and its parental strain DSP10 were compared under iron depleted conditions. The mutant displayed a severe growth defect when grown in the presence of 120 μM iron chelator (Fig. 4A), and had only marginal growth when grown in the presence of 160 μM iron chelator (Fig.  4B). This suggests that SO3032, which is induced ~20 fold after 40 minutes of iron depletion, is required for surviving iron starvation.
To determine the significance of a protein degradation module for iron depletion, a gene encoding a hypothetical protein (SO2017) was selected for mutagenesis. This gene is located in the same operon as the heat shock protein (HtpG), and its amino terminus contains a thioredoxin domain that might participate in redox reactions.
Recently, it was shown to be required for bacterial survival of heat shock [34]. Although both the SO2017 mutant and its parental strain MR-1 grew almost indistinguisha-bly in the absence of iron chelator 2,2'-dipyridyl, the mutant failed to grow under iron depleted conditions after incubating for five days at 30°C (Fig. 4C). Therefore, lack of SO2017 prohibited the cells' ability to survive iron depletion.
Network analysis suggests that ten novel transcriptional factors might be involved in anaerobic energy metabolism. Among them, one gene (SO1415) was successfully inactivated from the MR-1 genome. The Fe(III) reduction rates of the SO1415 mutant and MR-1 were assessed by a ferrozine assay as described [35]. Less than half the amount of Fe(II) was produced in the mutant as compared to MR-1 after four hours of incubation (Fig. 5A), indicating that disruption of SO1415 impaired anaerobic Fe(III) reduction.
We also assessed possible growth deficiency of the SO1415 mutant under anaerobic conditions with 10 mM lactate as the electron donor, and one of the following non-metal electron acceptors: 10 mM thiosulfate, 3 mM fumarate, 2 mM TMAO, 1 mM nitrate or 3 mM DMSO. Significant growth deficiency was observed when thiosulfate or TMAO was supplied ( Fig. 5B and 5C). In contrast, only very mild growth deficiency was observed when DMSO, fumarate or nitrate was supplied, and no growth deficiency was detected under aerobic conditions (data not shown). Together, these results suggest that SO1415 is involved in regulating specific branches of anaerobic energy metabolism. (A) Growth of SO3032 mutant and its parent strain DSP10 were compared in the presence of 120 μM 2,2'-dipyridyl

Discussion
In this work we show that iron response in S. oneidensis is a rapid process at both molecular and physiological levels. The use of the membrane permeable iron chelator 2,2'dipyridyl rapidly sequesters both intra-and extra-cellular iron [7]. As a result, when cells were sampled after one minute of iron depletion, functional and regulatory genes of iron acquisition systems (HugA, TonB1, ExbB1, AlcA, SO4743 and SO2426) were induced greater than six fold. These genes appear to be direct targets of Fur as evidenced by the presence of "Fur boxes" in their promoters. Notably, Fur has a weak affinity for Fe(II) to form a repression complex [36], allowing for a rapid and sensitive response to a change in iron concentration.
The identification of an anaerobic energy metabolism module for iron experiments conducted under aerobic conditions is intriguing. A straightforward explanation would be that gradually limiting oxygen induces an anaerobic response when bacteria are grown in shake flask cultures. However, we show here that S. oneidensis continues to grow and cell density continues to increase during iron depletion, a condition which should further limit the concentration of oxygen and thus up-regulate the expression of genes as an anaerobic response. The expression of genes in the anaerobic energy metabolism module does not fit this model. Furthermore, a number of genes involved in aerobic response are induced by iron depletion (Table 4), while genes involved in anaerobic response and not cofactored with Fe(II) are not induced by iron depletion. These genes include UbiC, chorismate lyase involved in the initial steps of ubiquinone biosynthesis; Ppc, phosphoenolpyruvate carboxylase synthesizing oxaloacetate; Pta and AceK, enzymes converting pyruvate to acetate during fermentation; GlpK, glycerol kinase synthesizing glycerol-3-phosphate; Tdh, threonine dehydrogenase involved in the supply of reducing potential; and glutamate decarboxylase (SO1769), whose E. coli counterpart is highly induced by anaerobiosis [37]. Therefore, an alternative explanation is that an anaerobic energy metabolism module may function as iron-storage proteins to release previously sequestrated iron in their protein products during iron depletion and thereby elevate the intracellular free iron pool. The use of non-essential, ironcofactor proteins for iron homeostasis has been well documented in E. coli [10,15]. Notably, a microarray study in E. coli indicates that a large number of genes related to anaerobic energy metabolism (e.g., Hyb and Frd) function in iron storage [10]. Such a mechanism seems to be present in S. oneidensis as well.
Ten transcriptional regulators were grouped within the module of anaerobic energy metabolism, suggesting that they may be involved in this process. We were able to generate a mutant of one of those genes (SO1415) and to experimentally verify its role in anaerobic energy metabolism. Genome analyses reveal that S. oneidensis has a large repertoire of transcriptional regulators, e.g., 88 two-component regulatory system proteins that enable the organism to adapt to a diversity of environmental conditions [12]. Nevertheless, most of the transcriptional regulators remain unstudied. The grouping of transcriptional regulators in anaerobic energy metabolism is an exciting finding; understanding this process is crucial to the potential utilization of Shewanella species to remediate U.S. Department of Energy's uranium-contaminated sites. In this regard, the rest of the transcriptional factors identified for anaerobic energy metabolism besides SO1415 are worthy of further investigation.
Bioinformatics analyses suggest that genes in the anaerobic energy metabolism module may be directly regulated by Crp. In E. coli, Crp modulates different biological processes and responds to glucose levels as a global transcriptional factor [38]. However, in S. oneidensis, Crp plays a critical role in anaerobic energy metabolism [31]. The identification of a Crp-binding site in this module provides a reasonable explanation for the function of Crp in multiple branches of anaerobic energy metabolism.
Transcriptomics and genetic studies suggest that protein degradation is involved in iron response. Lack of iron as a protein cofactor may impair the stability of a number of proteins. Induction of heat shock proteins may be necessary to process denatured or misfolded proteins. Intriguingly, no oxidative stress genes, such as Fe-superoxide dismutase (SodB) and genes in the SOS pathway (e.g., RecA, RpoD, RpoH, LexA, SsB, UmuC and UmuD) [39], were induced when iron was repleted. This is surprising since the excess of external iron is expected to provoke oxidative stress. It is possible that the concentration of iron used in this study is not sufficient to induce oxidative stress, or that S. oneidensis had yet not responded to oxidative stress during the period of time examined. Another possibility is that S. oneidensis employs novel pathways for oxidative stress response.
In E. coli, the TCA enzymes SdhA and AcnA are controlled by the regulatory cascade of Fur and the small RNA RyhB. Consequently, they are repressed under iron-depleted conditions [10,15]. This is not observed in S. oneidensis (Table 1). Indeed, we found that the expression of SdhA and AcnA in S. oneidensis was regulated neither in the fur mutant [2,3,25], nor in a strain that over-expresses RyhB the abolishment of gene expression in fur mutants [3,25]. This mechanism is well conserved among γ-proteobacteria [1,9]. In contrast, it appeared that the regulation of anaerobic energy metabolism and protein degradation modules was largely Fur-independent, as very similar sets of genes also were identified when a fur mutant was exposed to iron depletion and repletion conditions [25]. This implies that transcriptional regulators other than Fur are also essential for iron response. Indeed, Fur-independent regulation of gene expression by iron has been observed in E. coli, V. cholera, and H. Pylori [1,7,8]. Nevertheless, we could not completely rule out the possibility of an indirect effect of Fur on anaerobic energy metabolism and protein degradation. It was notable that an earlier study showed that Crp was differentially expressed in a fur mutant [3], despite the fact that there was no obvious "Fur box" upstream of the Crp ORF.

Conclusion
This report provides a glimpse into the physiological and molecular events of iron response in S. oneidensis. In addition to the Fur-regulated iron acquisition systems, numerous other genes are affected by changes of iron concentration. It will be interesting to further determine the mechanism of their involvement and regulation in iron response in future studies.

Bacterial strains and plasmids
Wild-type S. oneidensis MR-1 was used for physiology and transcriptomic studies. To generate an SO2017 deletion mutant from MR-1, the majority of the ORF was removed using PCR amplification with the primers A1 (5'AGCCT-GTGAGCTCACGGG), A2 (5'TGTTTAAACTTAGTGGAT-GGGGGTTAG ATCGAGGATATT), B1 (5'CCCATCCAC TAAGTTTAAACAGTTTGGCAAACCAAT ATC) and B2 (5'ACAATCGAGCTCTGCGAT), and a second cross-over PCR amplification with A1 and B2 using the mixed amplified fragments as templates was performed. The resulting product was cloned into the suicide plasmid pDS3.0 [3] and transformed into the E. coli WM3064 strain prior to conjugal transfer into MR-1. Correct in-frame deletion was verified by DNA sequencing of the region surrounding the DNA recombinant site. To generate a SO1415 mutant from MR-1, an internal fragment of SO1415 was amplified by PCR using primers (5'-TCCTTCGGACTC-CCTGT; 3'-CCATCAGGTTTGCTAAATGT) and cloned into the suicide plasmid pKNOCK-Km r with E. coli WM3064 as the host. After introducing the plasmid into MR-1 by conjugation, PCR was employed to amplify the genomic region surrounding the SO1415 locus, and subsequent sequencing confirmed the insertional disruption of SO1415. SO3032 and its parental strain DSP10 were described previously [40].
Physiological studies and ferrozine assay MR-1 was grown to mid-log phase (OD 600 = ~0.6) and diluted 1:100 into 300 μl fresh LB liquid medium. A Type FP-1100-C Bioscreen C machine (Thermo Labsystems) was used to measure growth every 30 min. All physiological studies were done in triplicates so that the average and standard deviation could be calculated. A range of concentrations of the iron chelator 2,2'-dipyridyl (80 μM, 160 μM, 240 μM, and 320 μM) was prepared in water. To measure the Fe(III) reduction rate, cells were grown anaerobically to mid-log phase in 10 ml fresh LB medium supplemented with 10 mM fumarate and 10 mM lactate. The cells were then spun down, washed with anaerobically prepared LB medium, and divided into aliquots of 5 × 10 7 cells that were transferred into 5 ml LB medium supplemented with 10 mM lactate and 10 mM Fe(III) dioxide in a Coy anaerobic chamber. The ferrozine assay was performed as previously described [35]. Uninoculated medium containing 10 mM Fe(III) served as the abiotic control.

RNA preparation
Four biological replicates of MR-1 were grown to the midlog phase (OD 600 = 0.6) in 100 ml LB medium in 500 ml shake flasks. 5 ml samples were collected at time 0, and then at 1, 5, 10, 20, 40, and 60 min after adding 2,2'-dipyridyl to attain a final concentration of 160 μM. Thereafter, ferrous sulfate was added, and cells were collected at 1, 5, 10, 20, 40, and 60 min. Cultures were vigorous shaked at 250-300 rpm to improve aeration. Cultures were sampled by centrifuging 6 ml cells at 14 krpm for one minute. After removing the supernatant, the cells were snap frozen in liquid nitrogen and stored at -80°C. Total RNA was extracted using Trizol Reagent (Invitrogen) as previously described [3]. RNA samples were treated with RNase-free DNase I (Ambion) to digest residual chromosomal DNA and then purified with RNeasy Kit (Qiagen) prior to spectrophotometric quantification at 260 and 280 nm.

Microarray hybridization, scanning and quantification
To allow for comparison of any pairs of samples, MR-1 genomic DNA was used as a common reference in each microarray experiment. This strategy has been successfully employed in several studies [41][42][43]. In brief, cDNA was produced in a first-strand reverse transcription (RT) reaction with random hexamer primers (Invitrogen) and labeled with Cy5 dUTP (Amersham Biosciences) by direct labeling. S. oneidensis MR-1 genomic DNA (gDNA) was amplified by Klenow (Invitrogen), and Cy3 dUTP (Amersham Biosciences) was incorporated into the product. Fluorescein-labeled probes were purified using the Qiaquick PCR purification kit (Qiagen).
A ScanArray Express Microarray Scanner (PerkinElmer) was used to scan the slides. Fluorescence and background intensity were quantified using ImaGene 6 software (Bio-Discovery). All spots with signal-to-background ratios of less than 3 were discarded.
Principles and details of Random Matrix Theory (RMT) based algorithms have been described in [34,45,46]. Application of RMT to large-scale biological data sets allows for grouping of functionally related genes and functional prediction of unknown genes. Compared to other network methods, RMT is unique in that it adopts two mathematical models to define the optimal threshold for removing microarray noise and hence produces an unsupervised network. The procedure to construct a gene co-expression network from the temporal gene expression profiles is as follows. The first step was to construct a gene expression correlation matrix M, whose elements are pairwise Pearson correlation coefficients (r) of the experimental results. Next, a series of correlation matrices were derived from M using different cutoff values. If the absolute value of an element in M is less than the selected cutoff, the element is set to 0. Eigenvalues of the derived matrices were obtained by direct diagonalization of the matrix. Standard spectral unfolding techniques [47] were applied to have a constant density of eigenvalues and thus to determine the nearest neighbor spacing distribution P(s), which describes the fluctuation of eigenvalues of the correlation matrices.
We used the chi square test to determine two critical threshold values: r l at which P(s) starts to deviate from the Gaussian orthogonal ensemble at a confidence level of p = 0.001 and r h at which P(s) follows the Poisson distribution at a confidence level of p = 0.001. The critical point r h was used as the threshold for constructing the gene coexpression network. Applying the RMT method to the microarray data revealed a Pearson correlation coefficient of 0.91 as the minimum threshold to construct the gene co-expression network. Since gene co-expression networks are hierarchical [48], thresholds higher than 0.91 were used to identify additional functional modules, resulting in the twelve modules shown in file. The Pajek software [49] was used to visualize the gene co-expression network.

Microarray data accession
The microarray data discussed in this publication have been deposited in NCBI's Gene Expression Omnibus and are accessible through GEO accession number GSE15334