Skip to main content

Using gene expression data and network topology to detect substantial pathways, clusters and switches during oxygen deprivation of Escherichia coli



Biochemical investigations over the last decades have elucidated an increasingly complete image of the cellular metabolism. To derive a systems view for the regulation of the metabolism when cells adapt to environmental changes, whole genome gene expression profiles can be analysed. Moreover, utilising a network topology based on gene relationships may facilitate interpreting this vast amount of information, and extracting significant patterns within the networks.


Interpreting expression levels as pixels with grey value intensities and network topology as relationships between pixels, allows for an image-like representation of cellular metabolism. While the topology of a regular image is a lattice grid, biological networks demonstrate scale-free architecture and thus advanced image processing methods such as wavelet transforms cannot directly be applied. In the study reported here, one-dimensional enzyme-enzyme pairs were tracked to reveal sub-graphs of a biological interaction network which showed significant adaptations to a changing environment. As a case study, the response of the hetero-fermentative bacterium E. coli to oxygen deprivation was investigated. With our novel method, we detected, as expected, an up-regulation in the pathways of hexose nutrients up-take and metabolism and formate fermentation. Furthermore, our approach revealed a down-regulation in iron processing as well as the up-regulation of the histidine biosynthesis pathway. The latter may reflect an adaptive response of E. coli against an increasingly acidic environment due to the excretion of acidic products during anaerobic growth in a batch culture.


Based on microarray expression profiling data of prokaryotic cells exposed to fundamental treatment changes, our novel technique proved to extract system changes for a rather broad spectrum of the biochemical network.


Over the last decades our understanding of cellular metabolism has increased considerably [1], in particular for less complex organisms such as Escherichia coli [24]. The gained knowledge includes cellular adaptation programs that respond to changing environmental conditions such as nutrient excess and starvation [5]. Current microarray technology allows for the investigation of all genes of an organism under various conditions, resulting in the generation of a massive amount of expression data. One of the greatest challenge we are faced with is to then analyse the data as a whole and extract the meaningful relationships among specific genes. Standard methods such as SAM [6] or machine learning algorithms [7] are able to detect patterns in gene expression data, distinguishing between different states of a cell. However, the above methods for classification and pattern discovery do not consider interactions between different genes and their corresponding proteins. Functional relationships between genes can be assembled by e.g. regulatory, signal transduction and metabolic networks. Gardner and co-workers used gene expression microarray data to infer a regulatory network for E. coli [8]. They developed a linear model and effectively reduced the number of parameters by assuming a sparse regulatory network. Finally, they verified their inferred regulatory network on a smaller subset, i.e. the regulation of the SOS pathway. In a recent study, a large compendium of gene expression microarray data for E. coli was analysed using an information theoretical approach revealing new regulatory interactions [9]. When analysing a metabolic network, every enzyme can be represented by its corresponding gene. For sets of genes, pathway scores have been calculated improving the sensitivity to detect crucial enzymatic pathways when taking network distances for enzyme pairs into account [10]. Transcription data and the topological information derived from the metabolic network was connected by calculating Z-scores of highly correlated sub-networks [11]. Genes with common biological processes or functions were grouped by their gene ontology terms [12] and gene set enrichment tests performed on these groupings [13]. Additionally, gene set enrichments were tested by their common pathways in the corresponding networks [14, 15]. However, these approaches do not take into account direct interactions within the network. In contrast, a Potts-spin clustering algorithm on metabolic networks was developed depending on direct nearest-neighbour relationships. It was applied yielding sub-graphs stimulated by environmental conditions [16]. Furthermore, common gene expression levels of neighbouring nodes in a metabolic network were calculated by averaging over all neighbours of a gene and revealed several interesting regulated pathways for the human immune system [17]. Rapaport and co-workers extracted gene expression patterns of neighbouring genes in the network yielding good classification of the profiled samples by calculating Fourier transformations and rejecting high frequency signals [18]. However, these approaches did not consider switch like behaviours of neighbouring genes. To detect common and contrasting tendencies, an image-like representation of the cellular metabolism can be used by interpreting expression levels as pixel intensities with grey values and the network topology as relationships between pixels. Image processing methods may then be applied to extract crucial features from such an image. Wavelet transforms are such an image processing method and were applied for texture classification [19], for feature generation to automatically classify microscope images [20] and large-scale functional genetic screens [21]. Without taking any network information into account, wavelet transforms have been used together with other image processing methods for analysing microarray data [22, 23], in particular the Haar wavelet power spectrum for feature selection [24]. The application potential of this powerful technology to analyse biological networks is clear, yet challenging. While the underlying topology of an ordinary image consists of a lattice grid, biological networks have a rather scale-free architecture [25]. We recently reported one approach that applied image processing methods on the two-dimensional and therefore image-like adjacency matrix of the network [26]. In the present study we expand upon this method using the original architecture of the metabolic network. We analysed gene expression changes for each pair of neighbouring nodes combining their values additive (common response) and subtractive (opposing response, switch like behaviour). In a second step all combined nodes with a common response were again combined to yield significant clusters of co-expression. Such a simple approach allowed the analysis of the cellular stress response, not only for highly connected regions of the network but also for linear chains as well as the identification of specific switches. We analysed gene expression changes of E. coli during oxygen deprivation. With this technique we were able to detect the expected substantial regulatory adaptation programs, including up-regulated formate fermentation, mixed acid fermentation, metabolisms of hexoses and down regulation of the respiratory TCA cycle (see Figure 1). Furthermore, our technique revealed a down-regulation of the iron processing metabolism due to reduced oxidative stress during oxygen deprivation. The revealed up-regulation of the histidine biosynthesis pathway may constitute the adaptive response of E. coli to an acidic environment due to the excretion of acidic products during anaerobic growth in a batch culture.

Figure 1

Overview of the found significantly regulated metabolic pathways during oxygen deprivation (green: up-regulated, red: down-regulated, blue: metabolites). For more details, see: Figure 4 (yellow cross-hatched), Figure 5 (blue hatched) and Figure 6 and 7 (red and light blue boxes, respectively). Note that in this Figure, the metabolic pathways of Figure 7 are represented by two boxes. This is due to the unspecific hub-like nature of L-glutamine (see Conclusions).

Results and Discussion

Testing the method with simulated data

Setting up the network and calculating the simulated expression data

To test our method with simulated data on a simplified model network, we constructed a regular grid of 30 × 40 artificial reactions (workflow see Figure 2). On this simulated image-like metabolic network we randomly selected pathways of connected reactions with lengths 7, 10 and 24. These lengths corresponded to an expected length of a biological pathway (7, 10) and to the most frequent path length of the shortest paths between all pairs of nodes in the regular grid, respectively. 100 runs were performed generating 44 experiments of simulated expression data with a ground level of 6, in rather good agreement with our normalised gene expression data. To this, a Gaussian noise of mean 0 and standard deviation 1 was added. Two classes were formed with 22 experiments each. In one class the reactions of the randomly chosen pathways were up-regulated by adding a constant level Δ to the random expression levels.

Figure 2

General workflow of the method. The metabolic network of E. coli was put up using the EcoCyc database. Gene expression data was mapped onto the reactions of the network resulting in an image like representation (red boxes). Features were generated by using the Haar wavelet transformation on every connected reaction pair. The most discriminative features were identified by a t-test. Sub-graphs were built by connecting significant reaction pairs. Regions with identical regulation of more than four reactions were extracted (clusters). Reaction pairs with opposite regulation were identified as switches and were also extracted. The resulting pathways were analysed by literature scanning in-depth. Assembling the found pathways yielded an overall picture of the metabolic processes.

Performing the method on random data

In each run the expression data of all 44 samples (22 class 1 and 22 class 2) were mapped on the nodes (reactions) of the simulated metabolic network. Features were generated by applying the one dimensional Haar-wavelet transform onto each pair of neighbouring nodes. This yielded 9320 features for every sample. A t-test was applied for every feature to rank the features with respect to their discriminating property while correcting p-values for multiple testing (Bonferroni) [27]. Every feature for all reaction-pairs was ranked according to its p-value. The p-value cut-off was set to 0.01. Reactions were regarded as up-regulated if the corresponding simulated genes were significantly differentially expressed (p-value of a t-test ≤ 0.05) and not significantly differentially expressed otherwise. Not differentially expressed end-nodes were discarded. We compared our technique to a standard method.

Comparison to a standard method

A standard Students t-test was applied on the simulated expression data without taking any network information into account. For both methods true positives, false positives, false negatives and true negatives were calculated. To investigate a broader spectrum for the precision and sensitivity of our technique, the validation was performed with a variety of added constants (Δ = 2, 4, 6). Our technique decreased the number of false positives significantly (Figure 3). In a step further we investigated how our technique performed on a biological network, choosing the metabolic network of E. coli, constructed as described in Methods. Out of this network we selected randomly pathways of lengths 5, 7, 10 and performed the same method as described above for different constants (Δ = 2, 4, 6). We obtained a similar superior performance of our approach. The number of false positives was reduced nearly threefold, while the detection power of true positives was identical (Results not shown).

Figure 3

Validation of the method on a regular grid consisting of 40 × 30 reactions (pixel). Random gene expression data was generated and mapped onto the nodes of the grid. The 44 samples were divided into two classes differing only significantly in the reactions of three randomly chosen pathways (red arrows). Up-regulation of these reactions in one class was achieved by adding a constant value Δ to their expression levels. Our technique revealed significantly less false positives (FP) than the standard t-test for all chosen values of Δ. The last row shows the desired outcome after 100 runs (TP: true positives, FP: false positives, TN: true positives, FN: false negatives).

The metabolism of E. coli under oxygen deprivation

The general workflow was briefly as follows:

  • Establishing the metabolic network using the EcoCyc database,

  • Mapping gene expression data onto the nodes of the network,

  • Generating feature modules using the Haar wavelet transformations,

  • Statistical testing of the feature modules,

  • Clustering of significant reaction pairs,

  • Analysing found clusters and switches in depth, and

  • Fusing of the results to receive an overall map of metabolic changes.

The expression data of all 43 samples (21 aerobic and 22 anaerobic) from the study of Covert et al. [28] was mapped onto the reactions of the metabolic network. Features were generated as described above using the Haar-wavelet transform yielding 6890 features for each sample. Discriminating features were identified via ranking of p-values from a t-test. Calculated p-values were corrected with a multiple t-test correction for possible mutant influences (see Methods) and for multiple testing (Bonferroni) [27]. The p-value cut-off was again set to 0.01 resulting in 660 significantly discriminating features. All significant reaction-pairs were extracted and connected yielding sub-graphs. In total, five such connected sub-graphs were identified consisting of 165 reactions. Reactions were regarded as up-regulated if the corresponding genes were significantly over-expressed under anaerobic conditions (p-value ≤ 0.05 of a mutant corrected t-test, see Methods), as down-regulated if significantly under-expressed and not significantly differentially expressed otherwise. Neighbouring, connected nodes that showed identical regulation (up or down) were grouped together to simplify interpretation. We refer to these groups as "clusters" in the following. Not differentially expressed end-nodes in a cluster were discarded. The resulting 10 clusters containing at least five reactions were interpreted in more detail and grouped according to their functional role (Table 1, supplement 1 contains the corresponding EcoCyc reaction-ids). Furthermore, pairs of reactions with significantly opposing regulatory behaviour were defined as switches. All significant switches were extracted (p-value ≤ 0.01). 64 such switches were identified. The first 20 switches are discussed in detail (Table 2 shows the first 20 switches, supplement 2 provides all 64 switches). An overview of the extracted pathways is given in the next paragraph.

Table 1 Extracted network clusters.
Table 2 Extracted switches in the network. Significantly differentially expressed pairs of reactions (p-value ≤ 0.01). The first 20 switches are shown here and described in detail in the text.

Main functional findings

The metabolic network of E. coli underwent substantial changes in regulation, when adapting to the environmental change from oxygen rich to deprived conditions (Figure 1). Due to limited oxygen, glycolysis and the fructose/mannose metabolism was up-regulated securing energy production under anaerobic conditions (Figure 5). Furthermore, pyruvate metabolism, formate fermentation and mixed acid fermentation were also up-regulated fermenting the products of the glycolysis (Figure 4). In contrast, iron processing and oxidative stress responses were down-regulated as oxidative stress was reduced (Figure 6). As expected, the aerobic part of the TCA-cycle was down-regulated. The need to generate essential compounds and amino acids was indicated by an elevated level of the aspartate metabolism (Figure 7). An indirect effect of the oxygen rich to oxygen deprived conditions was the up-regulation of the histidine biosynthesis (Figure 7). Histidine may function as a buffer for produced acids accumulating in the batch culture. In the following, these findings are described in detail.

Figure 4

Fermentation of formate was up-regulated processing pyruvate into formate via pyruvate lyase. Pyruvate is degraded to formic acid (formate), which then is either expelled (via transporters), or further degraded into H2 and CO2 by the formate hydrogenlyase complex (for more details see text). Reactions are symbolised by squares, metabolites by circles. Green (red) squares indicate significant up (down)-regulation (p-value ≤ 0.05) under anaerobic conditions.

Figure 5

Metabolisms of hexose nutrients under anaerobic conditions. Fructose and mannose metabolism were up-regulated indicating a higher glucose processing. For box colours see Figure 4.

Figure 6

Iron processing in an anaerobic environment. Iron is scavenged by E. coli using enterobactin, whose biosynthesis (blue bordered nodes) was down-regulated. For box colours see Figure 4.

Figure 7

During anaerobic growth E. coli performed mixed acid fermentation, resulting in a more acidic environment. The histidine biosynthesis (blue bordered nodes and light blue box) was up-regulated for buffering (see text). For box colours see Figure 4. PPR: phosphoribosyl.

Functional description of the extracted clusters

Pyruvate processing, formate fermentation, anaerobic respiration and anaerobic synthesis of deoxyribonucleosides

Two clusters belonged to this sub-group (Table 1). The first cluster (Figure 4) consisted of nine reactions. Edges were due to the metabolites formate, fumarate and reduced menaquinone. Reactions connected via formate: Pyruvate formate lyase was up-regulated under anaerobic conditions to process pyruvate into formate (fermentation) [29]. Formate degradation into CO2 and H2 was supported by up-regulated formate hydrogen lyase [30, 31] and formate dehydrogenase. Formate release into the periplasm was facilitated by up-regulation of the corresponding transporter (EcoCyc-id: TRANS-RXN-1). Also up-regulated was 3,4-dihydroxy-2-butanone 4-phosphate synthase which functions as the first and rate limiting step in flavin mononucleotide (FMN) biosynthesis. It is notable that FMN functions as an electron mediator during anaerobic respiration [32]. Reactions connected by reduced menaquinone: Dimethyl sulfoxide reductase was up-regulated as it is needed in the anaerobic electron transport chain [33]. Also up-regulated was fumarate reductase which is used by E. coli during anaerobic growth [34]. In this reaction menaquinol acts as an electron acceptor, while fumarate can function as a terminal electron donor [35]. Further reactions in this cluster were connected with fumarate reductase via fumarate. Aspartate ammonia-lyase was up-regulated, to process aspartate into fumarate during anaerobic growth on glucose [36, 37]. The second cluster contained six up-regulated reactions which were connected by the metabolite S-adenosyl-L-methionine. Pyruvate formate-lyase (PFL)-deactivase was up-regulated. This enzyme is a catalyser for quenching and inactivating pyruvate formate-lyase and is expressed under anaerobic conditions to the same levels as pyruvate formate-lyase [38]. Corresponding to this, pyruvate formate-lyase activase was also up-regulated as it activates pyruvate formate-lyase in an anaerobic environment [39]. Pyruvate dehydrogenase requires the lipoate modification of complex subunits [40]. Lipoate synthase is needed for the biosynthesis of lipoic acid and is necessary for the anaerobic glycine cleavage system activity [41]. Note that reactions belonging to the normal glycine cleavage system were all down-regulated (see the cluster of one carbon units). Furthermore, two reactions, i.e. for the anaerobic nucleoside-triphosphate reductase activating system and the component ribonucleoside triphosphate reductase activase were up-regulated. These reactions needed to synthesise deoxyribonucleotides under anaerobic conditions [42, 43] are showing expression patterns analogous to pyruvate formate-lyase activase [44]. Also up-regulated was the anaerobic coproporphyrinogen III oxidase. In an anaerobic environment this reaction is necessary for the biosynthesis of hemes which are essential co-factors of the electron transport chain [45].

Processing of hexoses

Two clusters represented the processing of hexose nutrients during anaerobic growth (Table 1). The first cluster was formed by eight up-regulated reactions (Figure 5). Connections between reactions were due to metabolites D-fructose-6-phosphate, fructose-1,6-bisphosphate and β-D-glucose-6-phosphate. Two major pathways were involved in the cluster: glycolysis, and fructose/mannose metabolism. The Embden-Meyerhof pathway is used when switching from aerobic respiration to fermentation during growth under anaerobic conditions on minimal medium with glucose [46], yielding in a strong increase of glucose consumption [29, 46]. All reactions processing glucose down to fructose-1,6-biphosphate were up-regulated: Glucokinase converting glucose to glucose-6-phosphate and phosphoglucose isomerase transforming glucose-6-phosphate to fructose-6-phosphate/6-phosphofructokinase yielding fructose-1,6-bisphosphate. The increased conversion of D-fructose-6-phosphate to mannitol-1-phosphate generating the electron acceptor NAD+ normally produced in the Krebs cycle [47] explains the up-regulation of mannitol-1-phosphate 5-dehydrogenase. The higher amount of 1-phosphofructokinase is in agreement with previous findings [46]. The EIIMan transporter was up-regulated to increase the up-take of glucose. The second cluster of reactions processing hexose nutrients contained five up-regulated reactions which were connected due to the metabolites 2-keto-3-deoxy-6-phospho-gluconate, D-glyceraldehyde-3-phosphate and 1,3-diphosphateglycerate. Phosphoglycerate kinase, glyceraldehyde 3-phosphate dehydrogenase and triose phosphate isomerase are induced by anaerobiosis [48, 49]. Note that they are part of the glycolytic pathway of E. coli. Finally, phosphogluconate dehydratase and 2-keto-3-deoxy-6-phosphogluconate aldolase which are key enzymes of the Entner-Doudoroff pathway, were up-regulated, further demonstrating anaerobic glucose metabolism [50].

Iron processing

Two clusters represented the processing of iron in an anaerobic environment (Table 1). The first cluster contained eight down-regulated reactions. From these, seven belonged to the complete biosynthesis pathway of enterobactin which is used by E. coli to scavenge iron, starting with isochorismate synthase and ending at the aryl carrier protein [51] (see Figure 6). Enterobactin biosynthesis is repressed under anaerobic conditions as it is used for aerobic iron transport [52]. Directly connected to the enterobactin pathway is enterochelin esterase. Enterochelin esterase uses enterobactin as an educt [53]. As biosynthesis of enterobactin was down-regulated, the down-regulation of enterochelin esterase is explained by the lower availability of its educts. The second cluster for iron processing contained five down-regulated reactions. Metabolites connecting the reactions were L-alanine and L-cysteine. The majority of the reactions are involved in Fe-S cluster biogenesis. The most connected node was cysteine desulfurase. This reaction assembles Fe-S complexes into Fe-S proteins to repair them when damaged during oxidative stress [54]. Under anaerobic conditions damage by oxidative stress is negligible explaining the down-regulation of cysteine desulfurase [55], whereas up-regulation as an oxidative stress response has been reported under aerobic conditions [56]. Directly linked to cysteine desulfurase was the thiamin (thiazole moiety) biosynthesis protein, which is a catalyser transferring sulfur from cysteine to the ThiS protein. It was down-regulated because during anaerobic growth a lower level of thiamin is needed compared to aerobic conditions [54]. Furthermore, selenocysteine lyase was connected to cysteine desulfurase via alanine. Selenocysteine lyase seems to be regulated by IscR and to form an alternate pathway involved in Fe-S biogenesis under aerobic conditions. In an anaerobic environment this reaction is known to be down-regulated [55].

Acid response

One prominent cluster was formed by 18 up-regulated reactions. Ten of these represent the complete histidine biosynthesis pathway, beginning with ATP phosphoribosyltransferase and ending at histidinal dehydrogenase (Figure 7). When growing anaerobically on glucose, E. coli synthesises acids via mixed acid fermentation [46, 57] and histidine is used to buffer acidic milieu [58]. Another reaction in the cluster was CTP synthetase which was also up-regulated. Due to the down-regulation of NDP kinase the elevated levels of CTP synthetase are in agreement with previous findings [59] while the concrete functionality of this remains unclear. Furthermore, the cluster consisted of up-regulated reactions that needed or produced aspartate under anaerobic conditions. In yeast it was shown that the aspartate concentration is roughly 100 times higher in the cells under anaerobic conditions [60]. Generating aspartate may facilitate the biosynthesis of further amino acids and other essential compounds. E. coli has two known reactions catalysing the synthesis of asparagine, asparagine synthetase and aspartate-ammonia ligase. Both reactions were up-regulated during anaerobic growth, in agreement with previous findings [61]. The role of aspartate was further reinforced by the up-regulation of the GltP glutamate/aspartate DAACS transporter. Finally the cluster consisted of the starting points for the anaerobic de novo biosynthesis of NAD which were also up-regulated. This pathway uses L-aspartate to form NAD via L-aspartate oxidase and Quinolinate synthase [62, 63]. Although NAD may be constitutively produced, the up-regulation of both reactions makes sense, as it has been shown that Quinolinate synthetase is inactive when exposed to oxygen [62].

Nucleosides metabolisms

Two clusters indicating a change in the processing of nucleosides were found (Table 1). One cluster contained eight down-regulated reactions processing GTP, GDP and dGDP. GTP cyclohydrolase I was down-regulated to limit the biosynthesis of cost intensive folate and highly abundant formate under anaerobic conditions. Similarly, GDP kinase, dGDP kinase, GDP reductase, deoxyguanylate kinase and ribonucleoside-diphosphate reductase 2 were down-regulated, which may be due to reducing the metabolism of cost intensive purines. Similarly the down-regulation of GDP diphosphokinase and deoxyguanylate kinase can be explained. The second cluster consisted of six down-regulated reactions. Edges between reactions were due to metabolites UTP, UDP, UDP-galactose, α-D-glucose 1-phosphate or dTTP. The highest connected node was UDP kinase. Interestingly, this cluster compares to the cluster above showing down-regulated processing of cost intensive nucleosides.

One carbon units

Six down-regulated reactions formed a cluster showing the processing of one carbon units under anaerobic conditions. Metabolites connecting the reactions were glycine, H-protein-S-(aminomethyldihydrolipoyl)lysine and H-protein-(lipoyl)lysine. The central reaction was glycine dehydrogenase (decarboxylating) which together with aminomethyltransferase is part of the glycine cleavage system. Although it is reported that the glycine cleavage system is active under anaerobic conditions [41], the down-regulation stems from the fact that the corresponding reaction reduces NAD+ to NADH which is very costly due to the low availability of NAD+ [64]. The production of one-carbon units, for which the glycine cleavage system is used [65], was taken over by glycine hydroxymethyltransferase (see switches). Furthermore, lipoyl-protein ligase A was down-regulated to reduce pyruvate dehydrogenase and to increase pyruvate formate lyase activity [66]. As a response to oxidative stress the expression of glutathione synthetase increases [67]. In an anaerobic environment no oxidative stress is prevalent, explaining the down-regulation of glutathione synthetase. Glycine-tRNA synthetase was down-regulated which may be due to reduced growth under oxygen deprivation.

Functional description of significant switches

64 significant switches were found (p-value ≤ 0.01, Table 2, supplement 2). The first 20 are interpreted here. Switches belonging to the same metabolic process are described in common paragraphs.

Formate fermentation

Five switches (1, 3, 4, 8, 15) belonged to the fermentation of formate. The following reaction-pairs were up-regulated and down-regulated respectively: formate hydrogenase complex and formyltetrahydrofolate deformylase, FocA formate FNT transporter and formyltetrahydrofolate deformylase, formate hydrogenase complex and GTP cyclohydrolase I, formate dehydrogenase and formyltetrahydrofolate deformylase, FocA formate FNT transporter and GTP cyclohydrolase I. All switches formed an intersection between degradation and formation of formate. Due to the high abundance of formate in the cell under anaerobic conditions, the formation of new formate was down-regulated (see e.g. [68]), while the degradation of formate into CO2 and H2 and the transport of formate to the periplasm was up-regulated.

Mixed acid fermentation and anaerobic respiration

Four switches (2, 10, 11, 14) belonged to mixed acid fermentation and anaerobic respiration. The first of these switches was formed by up-regulated acetaldehyde dehydrogenase and down-regulated ethanolamine ammonia-lyase. The reactions were connected via the metabolite acetaldehyde. E. coli ferments glucose via acetyl-CoA to ethanol. The first step in this fermentation is catalysed by acetaldehyde dehydrogenase converting acetyl-CoA to acetaldehyde [69]. Ethanolamine ammonia-lyase catalyses the cleavage of ethanolamine to acetaldehyde and ammonia [70]. Ethanolamine can be used as a carbon and energy source under aerobic conditions [71], resulting in a down-regulation of the reaction under anaerobic conditions. In two switches fumarate reductase was up-regulated where 5'-phosphoribosyl-4-(N-succinocarboxamide)-5-aminoimidazole lyase and adenylosuccinate lyase were down-regulated. 5'-phosphoribosyl-4-(N-succinocarboxamide)-5-aminoimidazole lyase and adenylosuccinate lyase form a bifunctional enzyme. The metabolite connecting the differently regulated reactions was fumarate in both cases. Fumarate reductase was up-regulated as it is used by E. coli during anaerobic growth [34] with menaquinol acting as an electron acceptor, while fumarate functions as a terminal electron donor [35]. 5'-phosphoribosyl-4-(N-succinocarboxamide)-5-aminoimidazole lyase/adenylosuccinate lyase was down-regulated to reduce the biosynthesis of purines indicating the reduced growth under oxygen deprivation. Another switch consisted of up-regulated phosphoenolpyruvate carboxylase and down-regulated aspartate transaminase, connected by the metabolite oxaloacetate. Phosphoenolpyruvate carboxylase participates in mixed-acid fermentation of glucose [72] and is therefore up-regulated during anaerobic growth. Under anaerobic conditions the citrate cycle is shortened to a reductive branch. CoA and oxaloacetate is then further processed to succinyl-coenzyme A by two possible branches, either using aspartate transaminase or malate dehydrogenase [73]. With our finding it seems that the second branch is favoured.

One carbon units

Two switches (6, 7) were part of the metabolism of one-carbon units. The switches consisted of up-regulated serine hydroxymethyltransferase and down-regulated reactions of the glycine cleavage system (gcv system and glycine dehydrogenase (decarboxylating)). Both reactions produce 5,10-methylene-THF and are therefore major contributors of one-carbon units in E. coli [65, 74, 75]. The switch found here indicates that under anaerobic conditions one-carbon units are more produced by serine hydroxymethyltransferase than in an aerobic environment. The glycine cleavage system reduces NAD+ to NADH. This is a costly reaction as NAD+ is only available in small quantities [64]. Therefore, the glycine cleavage system is down-regulated resulting in an up-regulation of serine hydroxymethyltransferase to compensate for the loss in one-carbon units.

Processing of hexoses

Two switches (9, 13) belonged to the processing of gluconate and glyoxylate. Both up-regulated nodes, 2-ketoaldonate reductase and 2-keto-4-hydroxyglutarate aldolase, participate in the intracellular regulation of glyoxylate levels [76, 77]. Gluconokinase converts gluconate to 6-phosphogluconate which then can enter the Entner-Doudoroff or the pentose phosphate pathway [77] while 2-keto-4-hydroxyglutarate aldolase forms a part of the Entner-Doudoroff pathway. Normally, the Entner-Doudoroff pathway is used if E. coli grows on gluconate. It exhibits basal levels of activity of this pathway if growing on glucose [78]. This is explained by the steady production of gluconate during glucose metabolism [78]. Under anaerobic conditions on glucose, the glucose metabolism is up-regulated (see Figure 5), which is followed by increased production of gluconate and an up-regulation of gluconate processing reactions. The down-regulated reactions, glyoxylate reductase and 2-ketoaldonate reductase respectively, use NADPH as the electron donor and cooperate with gluconate reductase [76] that, under aerobic conditions, brings glyoxylate into the tricarboxylic acid cycle [76]. Under anaerobic conditions this cycle is limited, resulting in the observed down-regulation.

Branched chain amino acids transporters

Three switches (16, 17, 18) were formed by branched chain amino acid transporters. Up-regulated were the BrnQ branched chain amino acid LIVCS transporters (EcoCyc-ids: TRANS-RXN-126, -126B, -126A). In contrast, branched chain amino acids ABC transporters (EcoCyc-ids: ABC-15-RXN, ABC-35-RXN, ABC-36-RXN) were down-regulated. The ABC transporters need costly ATP [79] resulting in a down-regulation under anaerobic conditions. To compensate for the loss of the high affinity ATP using transporters the low affinity branched chain amino acid LIVCS transporters were up-regulated [80].


An unexpected contrarily regulated reaction pair (switch 20) was formed by up-regulated phosphatidylglycerophosphate synthase and down-regulated CDP-diacylglycerol pyrophosphatase. E. coli uses phosphatidylglycerophosphate synthase to catalyse the biosynthesis of acidic phospholipids [81] synthesising phosphatidylglycerol. It plays a major role in translocation of e.g. trimethylamine N-oxide reductase [82], a reaction used for anaerobic respiration [83]. Unfortunately, not much functional knowledge exits about the down-regulated CDP-diacylglycerol pyrophosphatase. Switch 12 consisted of up-regulated CTP synthetase and down-regulated UDP-glucose-hexose-1-phosphate uridylyltransferase. The up-regulation of CTP-synthetase agrees to previous findings [59] although the reason for this remains to be investigated further. Two switches (5, 19) were inexplicable to us (see Conclusions).

Comparison to a standard method

We compared the list of genes extracted with our technique to a standard t-test which did not take any network information into account. A mutant-corrected t-test (see Methods) was run on the gene expression levels for the corresponding reactions. Table 3 shows the results for the first 40 highest ranking features. All except six reactions were also found by our technique. The top three reactions (1, 2, 3 of Table 3) are involved in fermentation of formate and were also found with our technique (Table 1, Figure 4). Our technique was capable detecting whole pathways that occurred in the list of our top ranking features. However, the standard method did not detect such pathways or sub-graphs (discussed in the text, see above) supporting our concept for identifying functionally relevant sub-graphs. The six reactions 10, 13, 15, 25, 29, 32 were not extracted by our technique. Five of these reactions were not found due to the network construction: Unspecific metabolites were deleted resulting in the deletion of reactions that catalyse unspecific substrates, such as pyruvate kinase, glutamate dehydrogenase (NADP+), NAD kinase, NADH oxidoreductase and RhtB homoserine Rht transporter. Putative reactions with undefined metabolites like N-acetyl-anhydromuramyl-L-alanine-amidase, were also not included into the studied metabolic network and could therefore not be identified.

Table 3 Discriminative reactions from a t-test. .


We applied simplified first- and second-order Haar-wavelet-transformations to select combined transcription levels of reaction-pairs. We chose the Haar wavelets as they enable connecting two discrete data points (reaction pairs in our case) in a straightforward way. Furthermore, we searched for common and opposing responses between combined gene expression data which matched well to the shape of the Haar wavelet filters. Through using this approach we gained substantial insight into the metabolic regulation of E. coli upon the transition from oxygen-rich to oxygen-deprived conditions. Such an approach complements to the original idea of DeRisi and co-workers to use microarray technology for discovering system changes. For example, they revealed changes in yeast metabolism during the diauxic shift [84]. In the study presented here, we discovered a broad spectrum of responses including direct responses to limited oxygen and changing buffer conditions. As a response to limited oxygen, we identified an up-regulation in glycolysis, other hexose metabolisms, mixed acid fermentation, formate fermentation and the metabolism of aspartate. In summary, we see two interesting implications for our study, (i) data analysis: the implementation of the Haar-wavelet technique on small pairs of nodes is well suited for revealing significant patterns in a cellular network; and (ii) functional: many pathways are regulated on a transcriptional level supporting the concept of hierarchical control analysis for micro-organisms [85, 86].

The formate fermentation showed an interesting switch like behaviour: for oxygen deprived conditions the degradation of formate was up-regulated while its cost-intensive production was down-regulated. Note that this may be more difficult to reveal when using smoothing techniques (as e.g. [17, 18]) and demonstrates the benefit of using wavelets. Furthermore, a decrease in the metabolism of iron was detected as a response to reduced oxygen availability. Interestingly, this agrees with Faith et al. who analysed a large compendium of 445 microarrays for E. coli including a variety of different oxygen conditions [9]. They showed that PdHR which regulates the central metabolism, is also involved in regulating the fec operon which encodes genes for iron transport. We discovered that the entire histidine biosynthesis pathway was up-regulated as a possible response to accumulation of acid products in batch culture [58]. However, essential sub-graphs were not only detected in an isolated form, but also in relation to connected pathways which depended on the same metabolites. E.g., the cluster containing the histidine biosynthesis pathway (Figure 7) also contained components for metabolism of aspartate and glutamine. In addition, the cluster of formate fermentation (Figure 4) included parts of the aspartate metabolism. This reflects the unspecific hub-like nature of key metabolites such as L-glutamine and aspartate connecting several pathways. Significant switches supported the yielded adaptation mechanisms of E. coli to changing oxygen abundance, as e.g., switches pertaining to the fermentation of formate and mixed acids.

In our previous study, we used the same microarray dataset and extracted discriminate patterns of highly connected regions in the network [26]. In comparison to the present study here, we got a good consistency of the extracted pathways (glycolysis, aspartate metabolism, formate fermentation, pyruvate metabolism). In the study presented here, we elucidated some new pathways, i.e. the histidine biosynthesis, enterobactin biosynthesis (oxidative stress response), the aerobic part of the TCA cycle, and hexoses and one-carbon-units processing. It is of note that histidine biosynthesis and the biosynthesis of enterobactin are linear chains in the network. In contrast to the previous method, such linear chains can be well recognised by the method we present here which couples pairs of nodes. However, the previous method recognised two interesting, highly-connected regions which were not indicated using our new method (the interface between glycolysis and NAD biosynthesis, and the biosynthesis of lysine, see [26]). Two observed switches remain explainable. One switch (switch 19) consisted of 3-hydroxy acid dehydrogenase and phosphoserine phosphatase. In the second case (switch 5) up-regulated 3-methyl-2-oxobutanoate hydroxymethyltransferase and down-regulated 2-dehydropantoate reductase directly followed the up-regulated reaction in the biosynthesis of pantothenate. These results may reveal an incomplete understanding of these metabolic components and the need for further experimental investigation.

On simulated data, the accuracy and precision was significantly better in comparison to the standard method. This allowed us to use a p-value of 0.01 and to focus on more significant changes. We compared our technique with a standard method extracting lists of discriminative genes from the expression data without taking gene relationships into account. We were able to detect all relevant reactions that could also be found by the standard method. In contrast, the standard method failed to reveal comprehensive functional pathways. However, for future studies a general method to validate the functionality of such a broad spectrum of newly revealed pathways remains to be developed. Nevertheless, our technique might be used for analysing signalling networks, e.g., to identify discriminative regulations in cancers with different prognosis, even though reaction and signalling levels might be less related to gene expression levels for higher organisms. Further methodological advances might also include the addition of protein post-transcriptional regulation and the application of more complex image processing methods.


Establishing the metabolic network

All metabolic reactions were extracted from the EcoCyc database (Version 10.0) [87]. A graph was established by defining neighbours of reactions: Two reactions were neighbours if a metabolite existed that was the product of one reaction and the substrate for the other. In this representation the nodes of the graph were the reactions while edges were defined by the metabolites. Metabolites were discarded that were highly connected and therefore pathway unspecific, such as water, oxygen, major coenzymes and prostethic groups. This approach resulted in a graph with 1196 nodes and 3650 edges.

Mapping gene expression data onto nodes of the network

Raw intensity values of gene expression data were collected from the work of Covert et al. in which mRNA levels of all open reading frames of E. coli using Affymetrix oligo microarrays were determined [28]. The data was downloaded from the ASAP database [88] and normalised with the variance normalisation method [89]. 43 hybridisations of the following samples were selected: strain K-12 MG 1655, wild-type, ΔarcA, ΔappY, Δfnr, ΔoxyR, ΔsoxS single mutants and the ΔarcAΔfnr double mutant. The mutated genes are key transcriptional regulators of the oxygen response [28]. They effect a major portion of all genes in E. coli. All gene expression experiments were done in triplicate under aerobic and anaerobic conditions, except for anaerobic wild-type which was repeated four times. The gene expression data of each data-set was mapped onto the corresponding reactions of the transcribed proteins. If a reaction was catalysed by a complex of proteins the minimal expression value of the genes involved was taken as the value of the corresponding complex. The expression data of all samples was mapped onto each network, yielding 43 different patterns for each graph.

Generating feature modules

To discover specific expression patterns and textures in the network we calculated features with the Haar-wavelet transformation consisting of gene expression combinations of neighbouring reaction-pairs. Haar-wavelet transformations add and subtract the values of neighbouring pairs of nodes and multiply them with a constant factor: Be r0, r1 the gene expression values of a pair of reactions, respectively. Applying the transform yields the feature modules f0, f1:

( f 1 f 0 ) = 1 2 ( 1 1 1 1 ) ( r 0 r 1 ) . MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaadaqadaqaauaabeqaceaaaeaacqWGMbGzdaWgaaWcbaGaeGymaedabeaaaOqaaiabdAgaMnaaBaaaleaacqaIWaamaeqaaaaaaOGaayjkaiaawMcaaiabg2da9maalaaabaGaeGymaedabaWaaOaaaeaacqaIYaGmaSqabaaaaOWaaeWaaeaafaqabeGacaaabaGaeGymaedabaGaeGymaedabaGaeGymaedabaGaeyOeI0IaeGymaedaaaGaayjkaiaawMcaamaabmaabaqbaeqabiqaaaqaaiabdkhaYnaaBaaaleaacqaIWaamaeqaaaGcbaGaemOCai3aaSbaaSqaaiabigdaXaqabaaaaaGccaGLOaGaayzkaaGaeiOla4caaa@4437@

See also [90] for more details. The Haar wavelet transform can be regarded as a low pass filter when performing the summation and a high pass filter when calculating the difference between neighbouring value pairs. Both filters were applied on all pairs of nodes connected by an edge resulting in calculated feature modules.

Statistical testing of the feature modules

All Haar-wavelet generated features were tested by a multiple t-test between aerobic and anaerobic conditions. To correct for potential influences coming from individual mutants, t-tests were performed for every constellation of samples excluding the sample of one particular mutant, respectively. The wild type sample was never excluded. From this outcome the worst (highest) p-value for each feature was selected. All p-values were corrected for multiple testing (Bonferroni, see [27]). Features were then ranked according to their p-value.

Clustering of significant reaction pairs

All significant feature modules were extracted (p-value ≤ 0.01). Sub-graphs were put up by connecting the found significant feature modules (reaction-pairs). This resulted in five larger sub-graphs. To facilitate the interpretation of the found sub-graphs, nodes with equal expression behaviour (up-, down-regulation) were grouped together. To reduce random fluctuations we focused only on larger patterns, i.e. clusters with a cluster size smaller than five were discarded. In total 10 clusters were extracted. Reaction-pairs having one up- and one down-regulated node were regarded as switches. They were extracted if their p-value was below 0.01.

Note, that our method yielding these clusters is based on two steps: generating feature modules and combining those with a common response yielding significant clusters of co-expression. The first step compares to a low and a high pass filter of the first order Haar-wavelet-transformation, respectively. The second step compares to a low pass filter of the second order Haar-wavelet-transform.

Analysing the found clusters and switches in-depth

All extracted clusters and switches were functionally characterised (see Results and Discussion). An in-depth analysis was performed by scanning the literature. Finally, the analysed clusters and switches were assembled yielding an overall map of the metabolic changes.


  1. 1.

    Berg JM, Tymoczko JL, Stryer L: Biochemistry. 5th edition. New York , Freeman; 2002.

    Google Scholar 

  2. 2.

    Caspi R, Foerster H, Fulcher CA, Hopkinson R, Ingraham J, Kaipa P, Krummenacker M, Paley S, Pick J, Rhee SY, Tissier C, Zhang P, Karp PD: MetaCyc: a multiorganism database of metabolic pathways and enzymes. Nucleic Acids Res 2006, 34(Database issue):D511–6. 10.1093/nar/gkj128

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  3. 3.

    Krieger CJ, Zhang P, Mueller LA, Wang A, Paley S, Arnaud M, Pick J, Rhee SY, Karp PD: MetaCyc: a multiorganism database of metabolic pathways and enzymes. Nucleic Acids Res 2004, 32(Database issue):D438–42. 10.1093/nar/gkh100

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  4. 4.

    Karp PD, Riley M, Paley SM, Pellegrini-Toole A: The MetaCyc Database. Nucleic Acids Res 2002, 30(1):59–61. 10.1093/nar/30.1.59

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  5. 5.

    Khodursky AB, Peter BJ, Cozzarelli NR, Botstein D, Brown PO, Yanofsky C: DNA microarray analysis of gene expression in response to physiological and genetic changes that affect tryptophan metabolism in Escherichia coli. Proc Natl Acad Sci U S A 2000, 97(22):12170–12175. 10.1073/pnas.220414297

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  6. 6.

    Tusher VG, Tibshirani R, Chu G: Significance analysis of microarrays applied to the ionizing radiation response. Proc Natl Acad Sci U S A 2001, 98(9):5116–5121. 10.1073/pnas.091062498

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  7. 7.

    Warnat P, Eils R, Brors B: Cross-platform analysis of cancer microarray data improves gene expression based classification of phenotypes. BMC Bioinformatics 2005, 6: 265. 10.1186/1471-2105-6-265

    PubMed Central  Article  PubMed  Google Scholar 

  8. 8.

    Gardner TS, di Bernardo D, Lorenz D, Collins JJ: Inferring genetic networks and identifying compound mode of action via expression profiling. Science 2003, 301(5629):102–105. 10.1126/science.1081900

    CAS  Article  PubMed  Google Scholar 

  9. 9.

    Faith JJ, Hayete B, Thaden JT, Mogno I, Wierzbowski J, Cottarel G, Kasif S, Collins JJ, Gardner TS: Large-Scale Mapping and Validation of Escherichia coli Transcriptional Regulation from a Compendium of Expression Profiles. PLoS Biol 2007, 5(1):e8. 10.1371/journal.pbio.0050008

    PubMed Central  Article  PubMed  Google Scholar 

  10. 10.

    Rahnenführer J, Domingues FS, Maydt J, Lengauer T: Calculating the statistical significance of changes in pathway activity from gene expression data. Stat Appl Genet Mol Biol 2004, 3(1):Article16.

    PubMed  Google Scholar 

  11. 11.

    Patil KR, Nielsen J: Uncovering transcriptional regulation of metabolism by using metabolic network topology. Proc Natl Acad Sci U S A 2005, 102(8):2685–2689. 10.1073/pnas.0406811102

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  12. 12.

    The Gene Ontology Consortium[]

  13. 13.

    Beissbarth T, Speed TP: GOstat: find statistically overrepresented Gene Ontologies within a group of genes. Bioinformatics (Oxford, England) 2004, 20(9):1464–1465. 10.1093/bioinformatics/bth088

    CAS  Article  Google Scholar 

  14. 14.

    Dahlquist KD, Salomonis N, Vranizan K, Lawlor SC, Conklin BR: GenMAPP, a new tool for viewing and analyzing microarray data on biological pathways. Nature genetics 2002, 31(1):19–20. 10.1038/ng0502-19

    CAS  Article  PubMed  Google Scholar 

  15. 15.

    Doniger SW, Salomonis N, Dahlquist KD, Vranizan K, Lawlor SC, Conklin BR: MAPPFinder: using Gene Ontology and GenMAPP to create a global gene-expression profile from microarray data. Genome biology 2003, 4(1):R7. 10.1186/gb-2003-4-1-r7

    PubMed Central  Article  PubMed  Google Scholar 

  16. 16.

    König R, Eils R: Gene expression analysis on biochemical networks using the Potts spin model. Bioinformatics 2004, 20(10):1500–1505. 10.1093/bioinformatics/bth109

    Article  PubMed  Google Scholar 

  17. 17.

    Nacu S, Critchley-Thorne R, Lee P, Holmes S: Gene Expression Network Analysis, and Applications to Immunology. Bioinformatics (Oxford, England) 2007.

    Google Scholar 

  18. 18.

    Rapaport F, Zinovyev A, Dutreix M, Barillot E, Vert JP: Classification of microarray data using gene networks. BMC bioinformatics 2007, 8: 35. 10.1186/1471-2105-8-35

    PubMed Central  Article  PubMed  Google Scholar 

  19. 19.

    Chang T, Kuo CCJ: Texture analysis and classification with tree-structured wavelet transform. IEEE Transactions on Image Processing 1993, 2(4):429–441. 10.1109/83.242353

    CAS  Article  PubMed  Google Scholar 

  20. 20.

    Conrad C, Erfle H, Warnat P, Daigle N, Lorch T, Ellenberg J, Pepperkok R, Eils R: Automatic identification of subcellular phenotypes on human cell arrays. Genome Res 2004, 14(6):1130–1136. 10.1101/gr.2383804

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  21. 21.

    Harder N, Mora-Bermúdez F, Godinez WJ, Ellenberg J, Eils R, Rohr K: Automated Analysis of the Mitotic Phases of Human Cells in 3D Fluorescence Microscopy Image Sequences: Copenhagen, DK. Volume 4190. Edited by: Larsen R, Nielsen M, Sporring J. Springer-Verlag; 2006:840–848.

    Google Scholar 

  22. 22.

    Klevecz RR: Dynamic architecture of the yeast cell cycle uncovered by wavelet decomposition of expression microarray data. Funct Integr Genomics 2000, 1(3):186–192. 10.1007/s101420000027

    CAS  Article  PubMed  Google Scholar 

  23. 23.

    Klevecz RR, Murray DB: Genome wide oscillations in expression. Wavelet analysis of time series data from yeast expression arrays uncovers the dynamic architecture of phenotype. Mol Biol Rep 2001, 28(2):73–82. 10.1023/A:1017909012215

    CAS  Article  PubMed  Google Scholar 

  24. 24.

    Subramani P, Sahu R, Verma S: Feature selection using Haar wavelet power spectrum. BMC Bioinformatics 2006, 7: 432. 10.1186/1471-2105-7-432

    PubMed Central  Article  PubMed  Google Scholar 

  25. 25.

    Jeong H, Tombor B, Albert R, Oltvai ZN, Barabasi AL: The large-scale organization of metabolic networks. Nature 2000, 407(6804):651–654. 10.1038/35036627

    CAS  Article  PubMed  Google Scholar 

  26. 26.

    König R, Schramm G, Oswald M, Seitz H, Sager S, Zapatka M, Reinelt G, Eils R: Discovering functional gene expression patterns in the metabolic network of Escherichia coli with wavelets transforms. BMC Bioinformatics 2006, 7: 119. 10.1186/1471-2105-7-119

    PubMed Central  Article  PubMed  Google Scholar 

  27. 27.

    Bonferroni CE: Il calcolo delle assicurazioni su gruppi di teste. In Studi in Onore del Professore Salvatore Ortu Carboni. Rome ; 1935:13–60.

    Google Scholar 

  28. 28.

    Covert MW, Knight EM, Reed JL, Herrgard MJ, Palsson BO: Integrating high-throughput and computational data elucidates bacterial networks. Nature 2004, 429(6987):92–96. 10.1038/nature02456

    CAS  Article  PubMed  Google Scholar 

  29. 29.

    Smith MW, Neidhardt FC: Proteins induced by anaerobiosis in Escherichia coli. J Bacteriol 1983, 154(1):336–343.

    PubMed Central  CAS  PubMed  Google Scholar 

  30. 30.

    Peck HD Jr., Gest H: Formic dehydrogenase and the hydrogenlyase enzyme complex in coli-aerogenes bacteria. J Bacteriol 1957, 73(6):706–721.

    PubMed Central  CAS  PubMed  Google Scholar 

  31. 31.

    Stewart V: Nitrate respiration in relation to facultative metabolism in enterobacteria. Microbiol Rev 1988, 52(2):190–232.

    PubMed Central  CAS  PubMed  Google Scholar 

  32. 32.

    Stams AJ, de Bok FA, Plugge CM, van Eekert MH, Dolfing J, Schraa G: Exocellular electron transfer in anaerobic microbial communities. Environ Microbiol 2006, 8(3):371–382. 10.1111/j.1462-2920.2006.00989.x

    CAS  Article  PubMed  Google Scholar 

  33. 33.

    Pruss BM, Liu X, Hendrickson W, Matsumura P: FlhD/FlhC-regulated promoters analyzed by gene array and lacZ gene fusions. FEMS Microbiol Lett 2001, 197(1):91–97.

    CAS  Article  PubMed  Google Scholar 

  34. 34.

    Iverson TM, Luna-Chavez C, Cecchini G, Rees DC: Structure of the Escherichia coli fumarate reductase respiratory complex. Science 1999, 284(5422):1961–1966. 10.1126/science.284.5422.1961

    CAS  Article  PubMed  Google Scholar 

  35. 35.

    Cecchini G, Schroder I, Gunsalus RP, Maklashina E: Succinate dehydrogenase and fumarate reductase from Escherichia coli. Biochim Biophys Acta 2002, 1553(1–2):140–157. 10.1016/S0005-2728(01)00238-9

    CAS  Article  PubMed  Google Scholar 

  36. 36.

    Falzone CJ, Karsten WE, Conley JD, Viola RE: L-aspartase from Escherichia coli: substrate specificity and role of divalent metal ions. Biochemistry 1988, 27(26):9089–9093. 10.1021/bi00426a004

    CAS  Article  PubMed  Google Scholar 

  37. 37.

    Guest JR, Roberts RE, Wilde RJ: Cloning of the aspartase gene (aspA) of Escherichia coli. J Gen Microbiol 1984, 130(5):1271–1278.

    CAS  PubMed  Google Scholar 

  38. 38.

    Kessler D, Leibrecht I, Knappe J: Pyruvate-formate-lyase-deactivase and acetyl-CoA reductase activities of Escherichia coli reside on a polymeric protein particle encoded by adhE. FEBS Lett 1991, 281(1–2):59–63. 10.1016/0014-5793(91)80358-A

    CAS  Article  PubMed  Google Scholar 

  39. 39.

    Hesslinger C, Fairhurst SA, Sawers G: Novel keto acid formate-lyase and propionate kinase enzymes are components of an anaerobic pathway in Escherichia coli that degrades L-threonine to propionate. Mol Microbiol 1998, 27(2):477–492. 10.1046/j.1365-2958.1998.00696.x

    CAS  Article  PubMed  Google Scholar 

  40. 40.

    Herbert AA, Guest JR: Lipoic acid content of Escherichia coli and other microorganisms. Arch Microbiol 1975, 106(3):259–266. 10.1007/BF00446532

    CAS  Article  PubMed  Google Scholar 

  41. 41.

    Reed KE, Cronan JE Jr.: Lipoic acid metabolism in Escherichia coli: sequencing and functional characterization of the lipA and lipB genes. J Bacteriol 1993, 175(5):1325–1336.

    PubMed Central  CAS  PubMed  Google Scholar 

  42. 42.

    Fontecave M, Mulliez E, Logan DT: Deoxyribonucleotide synthesis in anaerobic microorganisms: the class III ribonucleotide reductase. Prog Nucleic Acid Res Mol Biol 2002, 72: 95–127.

    CAS  Article  PubMed  Google Scholar 

  43. 43.

    Garriga X, Eliasson R, Torrents E, Jordan A, Barbe J, Gibert I, Reichard P: nrdD and nrdG genes are essential for strict anaerobic growth of Escherichia coli. Biochem Biophys Res Commun 1996, 229(1):189–192. 10.1006/bbrc.1996.1778

    CAS  Article  PubMed  Google Scholar 

  44. 44.

    Bianchi V, Reichard P, Eliasson R, Pontis E, Krook M, Jornvall H, Haggard-Ljungquist E: Escherichia coli ferredoxin NADP+ reductase: activation of E. coli anaerobic ribonucleotide reduction, cloning of the gene (fpr), and overexpression of the protein. J Bacteriol 1993, 175(6):1590–1595.

    PubMed Central  CAS  PubMed  Google Scholar 

  45. 45.

    Troup B, Hungerer C, Jahn D: Cloning and characterization of the Escherichia coli hemN gene encoding the oxygen-independent coproporphyrinogen III oxidase. J Bacteriol 1995, 177(11):3326–3331.

    PubMed Central  CAS  PubMed  Google Scholar 

  46. 46.

    Thomas AD, Doelle HW, Westwood AW, Gordon GL: Effect of oxygen on several enzymes involved in the aerobic and anaerobic utilization of glucose in Escherichia coli. J Bacteriol 1972, 112(3):1099–1105.

    PubMed Central  CAS  PubMed  Google Scholar 

  47. 47.

    Rosenberg H, Pearce SM, Hardy CM, Jacomb PA: Rapid turnover of mannitol-1-phosphate in Escherichia coli. J Bacteriol 1984, 158(1):63–68.

    PubMed Central  CAS  PubMed  Google Scholar 

  48. 48.

    Nellemann LJ, Holm F, Atlung T, Hansen FG: Cloning and characterization of the Escherichia coli phosphoglycerate kinase (pgk) gene. Gene 1989, 77(1):185–191. 10.1016/0378-1119(89)90373-9

    CAS  Article  PubMed  Google Scholar 

  49. 49.

    Gess B, Hofbauer KH, Deutzmann R, Kurtz A: Hypoxia up-regulates triosephosphate isomerase expression via a HIF-dependent pathway. Pflugers Arch 2004, 448(2):175–180. 10.1007/s00424-004-1241-1

    CAS  Article  PubMed  Google Scholar 

  50. 50.

    Pruss BM, Campbell JW, Van Dyk TK, Zhu C, Kogan Y, Matsumura P: FlhD/FlhC is a regulator of anaerobic respiration and the Entner-Doudoroff pathway through induction of the methyl-accepting chemotaxis protein Aer. J Bacteriol 2003, 185(2):534–543. 10.1128/JB.185.2.534-543.2003

    CAS  Article  PubMed  Google Scholar 

  51. 51.

    Buss K, Muller R, Dahm C, Gaitatzis N, Skrzypczak-Pietraszek E, Lohmann S, Gassen M, Leistner E: Clustering of isochorismate synthase genes menF and entC and channeling of isochorismate in Escherichia coli. Biochim Biophys Acta 2001, 1522(3):151–157.

    CAS  Article  PubMed  Google Scholar 

  52. 52.

    Kwon O, Hudspeth ME, Meganathan R: Anaerobic biosynthesis of enterobactin Escherichia coli: regulation of entC gene expression and evidence against its involvement in menaquinone (vitamin K2) biosynthesis. J Bacteriol 1996, 178(11):3252–3259.

    PubMed Central  CAS  PubMed  Google Scholar 

  53. 53.

    Greenwood KT, Luke RK: Enzymatic hydrolysis of enterochelin and its iron complex in Escherichia Coli K-12. Properties of enterochelin esterase. Biochim Biophys Acta 1978, 525(1):209–218.

    CAS  Article  PubMed  Google Scholar 

  54. 54.

    Schwartz CJ, Djaman O, Imlay JA, Kiley PJ: The cysteine desulfurase, IscS, has a major role in in vivo Fe-S cluster formation in Escherichia coli. Proc Natl Acad Sci U S A 2000, 97(16):9009–9014. 10.1073/pnas.160261497

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  55. 55.

    Giel JL, Rodionov D, Liu M, Blattner FR, Kiley PJ: IscR-dependent gene expression links iron-sulphur cluster assembly to the control of O-regulated genes in Escherichia coli. Mol Microbiol 2006, 60(4):1058–1075. 10.1111/j.1365-2958.2006.05160.x

    CAS  Article  PubMed  Google Scholar 

  56. 56.

    Zheng M, Wang X, Templeton LJ, Smulski DR, LaRossa RA, Storz G: DNA microarray-mediated transcriptional profiling of the Escherichia coli response to hydrogen peroxide. J Bacteriol 2001, 183(15):4562–4570. 10.1128/JB.183.15.4562-4570.2001

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  57. 57.

    Thayer DW: Characterization of some of the acidic products of glucose fermentation by a large-cell strain of Escherichia coli. Can J Microbiol 1970, 16(12):1243–1248.

    CAS  Article  PubMed  Google Scholar 

  58. 58.

    Gale EF, Epps HMR: The effect of the pH of the medium during growth on the enzymic activities of bacteria (Escherichia coli and Micrococcus lysodeikticus) and the biological significance of the changes produced. Biochemical journal 1942, 36(7–9):600–618.

    PubMed Central  PubMed  Google Scholar 

  59. 59.

    Shen R, Olcott MC, Kim J, Rajagopal I, Mathews CK: Escherichia coli nucleoside diphosphate kinase interactions with T4 phage proteins of deoxyribonucleotide synthesis and possible regulatory functions. J Biol Chem 2004, 279(31):32225–32232. 10.1074/jbc.M402750200

    CAS  Article  PubMed  Google Scholar 

  60. 60.

    Stephanopoulos G, Alper H, Moxley J: Exploiting biological complexity for strain improvement through systems biology. Nat Biotechnol 2004, 22(10):1261–1267. 10.1038/nbt1016

    CAS  Article  PubMed  Google Scholar 

  61. 61.

    Draczynska-Lusiak B, Brown OR: Asparagine synthetase: an oxidant-sensitive enzyme in Escherichia coli. Microbios 1994, 77(312):141–152.

    CAS  PubMed  Google Scholar 

  62. 62.

    Ollagnier-de Choudens S, Loiseau L, Sanakis Y, Barras F, Fontecave M: Quinolinate synthetase, an iron-sulfur enzyme in NAD biosynthesis. FEBS Lett 2005, 579(17):3737–3743. 10.1016/j.febslet.2005.05.065

    CAS  Article  PubMed  Google Scholar 

  63. 63.

    Foster JW, Moat AG: Nicotinamide adenine dinucleotide biosynthesis and pyridine nucleotide cycle metabolism in microbial systems. Microbiol Rev 1980, 44(1):83–105.

    PubMed Central  CAS  PubMed  Google Scholar 

  64. 64.

    Madigan TM, Martinko JM, Parker J: Biology of Microorganisms. 10th edition. Prentice Hall; 2003.

    Google Scholar 

  65. 65.

    Stauffer GV: Biosynthesis of Serine and Glycine. In Escherichia coli and Salmonella typhimurium Cellular and Molecular biology. Volume 1. Edited by: Neidhardt FC. American Society for Microbiology; 1987:412–418.

    Google Scholar 

  66. 66.

    Green DE, Morris TW, Green J, Cronan JE Jr., Guest JR: Purification and properties of the lipoate protein ligase of Escherichia coli. Biochem J 1995, 309 ( Pt 3): 853–862.

    CAS  Article  Google Scholar 

  67. 67.

    Izawa S, Inoue Y, Kimura A: Oxidative stress response in yeast: effect of glutathione on adaptation to hydrogen peroxide stress in Saccharomyces cerevisiae. FEBS Lett 1995, 368(1):73–76. 10.1016/0014-5793(95)00603-7

    CAS  Article  PubMed  Google Scholar 

  68. 68.

    Nagy PL, Marolewski A, Benkovic SJ, Zalkin H: Formyltetrahydrofolate hydrolase, a regulatory enzyme that functions to balance pools of tetrahydrofolate and one-carbon tetrahydrofolate adducts in Escherichia coli. J Bacteriol 1995, 177(5):1292–1298.

    PubMed Central  CAS  PubMed  Google Scholar 

  69. 69.

    Clark DP: The fermentation pathways of Escherichia coli. FEMS Microbiol Rev 1989, 5(3):223–234. 10.1016/0168-6445(89)90033-8

    CAS  PubMed  Google Scholar 

  70. 70.

    Roof DM, Roth JR: Functions required for vitamin B12-dependent ethanolamine utilization in Salmonella typhimurium. J Bacteriol 1989, 171(6):3316–3323.

    PubMed Central  CAS  PubMed  Google Scholar 

  71. 71.

    Kofoid E, Rappleye C, Stojiljkovic I, Roth J: The 17-gene ethanolamine (eut) operon of Salmonella typhimurium encodes five homologues of carboxysome shell proteins. J Bacteriol 1999, 181(17):5317–5329.

    PubMed Central  CAS  PubMed  Google Scholar 

  72. 72.

    Gokarn RR, Eiteman MA, Altman E: Metabolic analysis of Escherichia coli in the presence and absence of the carboxylating enzymes phosphoenolpyruvate carboxylase and pyruvate carboxylase. Appl Environ Microbiol 2000, 66(5):1844–1850. 10.1128/AEM.66.5.1844-1850.2000

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  73. 73.

    Goh EB, Bledsoe PJ, Chen LL, Gyaneshwar P, Stewart V, Igo MM: Hierarchical control of anaerobic gene expression in Escherichia coli K-12: the nitrate-responsive NarX-NarL regulatory system represses synthesis of the fumarate-responsive DcuS-DcuR regulatory system. J Bacteriol 2005, 187(14):4890–4899. 10.1128/JB.187.14.4890-4899.2005

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  74. 74.

    Meedel TH, Pizer LI: Regulation of one-carbon biosynthesis and utilization in Escherichia coli. J Bacteriol 1974, 118(3):905–910.

    PubMed Central  CAS  PubMed  Google Scholar 

  75. 75.

    Stauffer LT, Stauffer GV: Characterization of the gcv control region from Escherichia coli. J Bacteriol 1994, 176(20):6159–6164.

    PubMed Central  CAS  PubMed  Google Scholar 

  76. 76.

    Nunez MF, Pellicer MT, Badia J, Aguilar J, Baldoma L: Biochemical characterization of the 2-ketoacid reductases encoded by ycdW and yiaE genes in Escherichia coli. Biochem J 2001, 354(Pt 3):707–715. 10.1042/0264-6021:3540707

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  77. 77.

    Klemm P, Tong S, Nielsen H, Conway T: The gntP gene of Escherichia coli involved in gluconate uptake. J Bacteriol 1996, 178(1):61–67.

    PubMed Central  CAS  PubMed  Google Scholar 

  78. 78.

    Eisenberg RC, Dobrogosz WJ: Gluconate metabolism in Escherichia coli. J Bacteriol 1967, 93(3):941–949.

    PubMed Central  CAS  PubMed  Google Scholar 

  79. 79.

    Igarashi K, Kashiwagi K: Polyamine transport in bacteria and yeast. Biochem J 1999, 344 Pt 3: 633–642. 10.1042/0264-6021:3440633

    CAS  Article  PubMed  Google Scholar 

  80. 80.

    Anderson JJ, Oxender DL: Genetic separation of high- and low-affinity transport systems for branched-chain amino acids in Escherichia coli K-12. J Bacteriol 1978, 136(1):168–174.

    PubMed Central  CAS  PubMed  Google Scholar 

  81. 81.

    Gopalakrishnan AS, Chen YC, Temkin M, Dowhan W: Structure and expression of the gene locus encoding the phosphatidylglycerophosphate synthase of Escherichia coli. J Biol Chem 1986, 261(3):1329–1338.

    CAS  PubMed  Google Scholar 

  82. 82.

    Mikhaleva NI, Santini CL, Giordano G, Nesmeyanova MA, Wu LF: Requirement for phospholipids of the translocation of the trimethylamine N-oxide reductase through the Tat pathway in Escherichia coli. FEBS Lett 1999, 463(3):331–335. 10.1016/S0014-5793(99)01661-0

    CAS  Article  PubMed  Google Scholar 

  83. 83.

    Mejean V, Iobbi-Nivol C, Lepelletier M, Giordano G, Chippaux M, Pascal MC: TMAO anaerobic respiration in Escherichia coli: involvement of the tor operon. Mol Microbiol 1994, 11(6):1169–1179. 10.1111/j.1365-2958.1994.tb00393.x

    CAS  Article  PubMed  Google Scholar 

  84. 84.

    DeRisi JL, Iyer VR, Brown PO: Exploring the metabolic and genetic control of gene expression on a genomic scale. Science 1997, 278(5338):680–686. 10.1126/science.278.5338.680

    CAS  Article  PubMed  Google Scholar 

  85. 85.

    Rossell S, van der Weijden CC, Kruckeberg AL, Bakker BM, Westerhoff HV: Hierarchical and metabolic regulation of glucose influx in starved Saccharomyces cerevisiae. FEMS yeast research 2005, 5(6–7):611–619. 10.1016/j.femsyr.2004.11.003

    CAS  Article  PubMed  Google Scholar 

  86. 86.

    Rossell S, van der Weijden CC, Lindenbergh A, van Tuijl A, Francke C, Bakker BM, Westerhoff HV: Unraveling the complexity of flux regulation: a new method demonstrated for nutrient starvation in Saccharomyces cerevisiae. Proceedings of the National Academy of Sciences of the United States of America 2006, 103(7):2166–2171. 10.1073/pnas.0509831103

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  87. 87.

    Keseler IM, Collado-Vides J, Gama-Castro S, Ingraham J, Paley S, Paulsen IT, Peralta-Gil M, Karp PD: EcoCyc: a comprehensive database resource for Escherichia coli. Nucleic Acids Res 2005, 33(Database issue):D334–7. 10.1093/nar/gki108

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  88. 88.

    Glasner JD, Liss P, Plunkett G 3rd, Darling A, Prasad T, Rusch M, Byrnes A, Gilson M, Biehl B, Blattner FR, Perna NT: ASAP, a systematic annotation package for community analysis of genomes. Nucleic Acids Res 2003, 31(1):147–151. 10.1093/nar/gkg125

    PubMed Central  CAS  Article  PubMed  Google Scholar 

  89. 89.

    Huber W, von Heydebreck A, Sultmann H, Poustka A, Vingron M: Variance stabilization applied to microarray data calibration and to the quantification of differential expression. Bioinformatics 2002, 18 Suppl 1: S96–104.

    Article  PubMed  Google Scholar 

  90. 90.

    Theodoridis S, Koutroumbas K: Pattern Recognition. London , Academic Press; 1998.

    Google Scholar 

Download references


We gratefully thank David Jackson and Nathan Brady for stylistic corrections. We thank EcoCyc, Covert and his co-workers, and the ASAP team for making their data online available. The work was funded by the German National Genome Research Network (NGFN 01 GR 0450) and the BMBF (Viroquant).

Author information



Corresponding author

Correspondence to Roland Eils.

Additional information

Authors' contributions

GS and RK conceptualised and designed the method. GS analysed the data. RK, GS and MZ analysed and interpreted the data on its biological content. The manuscript was written by GS, MZ and RK. RE revised it critically. All authors read and approved the final manuscript.

Electronic supplementary material


Additional file 1: Supplement 1: Extracted network clusters with EcoCyc identifiers. The extracted network clusters with the identifiers from the EcoCyc database. (DOC 58 KB)


Additional file 2: Supplement 2: All significant switches in the network. List of all significant switches found in the network. (DOC 105 KB)

Authors’ original submitted files for images

Below are the links to the authors’ original submitted files for images.


Authors’ original file for figure 1


Authors’ original file for figure 2


Authors’ original file for figure 3


Authors’ original file for figure 4


Authors’ original file for figure 5


Authors’ original file for figure 6


Authors’ original file for figure 7

Rights and permissions

Open Access This article is published under license to BioMed Central Ltd. This is an Open Access article is distributed under the terms of the Creative Commons Attribution License ( ), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Reprints and Permissions

About this article

Cite this article

Schramm, G., Zapatka, M., Eils, R. et al. Using gene expression data and network topology to detect substantial pathways, clusters and switches during oxygen deprivation of Escherichia coli. BMC Bioinformatics 8, 149 (2007).

Download citation


  • Metabolic Network
  • Oxygen Deprivation
  • Enterobactin
  • Fumarate Reductase
  • Mixed Acid Fermentation