Volume 5 Supplement 2
22nd International Conference on Genome Informatics: Systems Biology
Framework for network modularization and Bayesian network analysis to investigate the perturbed metabolic network
 Hyun Uk Kim†^{1, 2},
 Tae Yong Kim†^{1, 2}Email author and
 Sang Yup Lee^{1, 2, 3}Email author
DOI: 10.1186/175205095S2S14
© Kim et al; licensee BioMed Central Ltd. 2011
Published: 14 December 2011
Abstract
Background
Genomescale metabolic network models have contributed to elucidating biological phenomena, and predicting gene targets to engineer for biotechnological applications. With their increasing importance, their precise network characterization has also been crucial for better understanding of the cellular physiology.
Results
We herein introduce a f ramework for network m odularization and B ayesian network analysis (FMB) to investigate organism’s metabolism under perturbation. FMB reveals direction of influences among metabolic modules, in which reactions with similar or positively correlated flux variation patterns are clustered, in response to specific perturbation using metabolic flux data. With metabolic flux data calculated by constraintsbased flux analysis under both control and perturbation conditions, FMB, in essence, reveals the effects of specific perturbations on the biological system through network modularization and Bayesian network analysis at metabolic modular level. As a demonstration, this framework was applied to the genetically perturbed Escherichia coli metabolism, which is a lpdA gene knockout mutant, using its genomescale metabolic network model.
Conclusions
After all, it provides alternative scenarios of metabolic flux distributions in response to the perturbation, which are complementary to the data obtained from conventionally available genomewide highthroughput techniques or metabolic flux analysis.
Background
Genomescale metabolic network models have proved to be useful in aiding biological and biotechnological research by providing largescale predications, and their scope of applications is still expanding [1]. They predict gene targets to manipulate in metabolic engineering for overproduction of industrially valuable chemicals [2, 3], or predict drug targets for drug discovery [4–7]. These metabolic network models are a coherently organized set of metabolites whose relationships are stoichiometrically defined and massbalanced through biochemical reactions [8]. They are often simulated with constraintsbased flux analysis that employs optimizationbased techniques with an objective function, typically maximizing biomass formation rate, and constraints that reflect biophysical conditions affecting the cellular physiology [9, 10]. Due to their importance, the number of currently reported genomescale metabolic networks is constantly increasing, spanning archaea, bacteria, and eukaryotes [1].
Reconstruction of metabolic network models has been accompanied by systematic characterization of their network properties for appropriate analysis of the prediction outcomes [11]. Some noteworthy constraintsbased approaches for elucidating correlations among intracellular reactions include flux coupling analysis [12], assessment of alternative sets of reactions for the optimal solutions [13], uniform random sampling [14], and elucidation of metabolic core reactions [15]. As an extension of these studies, correlations among metabolic modules that behave in a synchronized manner in the solution space allowed under the specific condition attracted attention for further investigations. For this, application of clustering (or modularization) and Bayesian network (BN) analysis to metabolic network models and their flux data was considered in this study in order to reveal other complementary aspects of the cell that are statistically and biologically important under the perturbation condition – modularlevel behaviour of metabolism.
Clustering has succeeded in revealing key components and their correlations, contributing to our understanding of biological systems through a systemic concept of module [16, 17]. Modularization of biological network, which is based on clustering algorithms, elucidates topological design of biological system, from which evolutionary and functional clues can be inferred [18, 19]. Metabolic network can also be clustered based on the characteristics of its metabolic fluxes, which are defined as the cellular phenotype derived from interplays of many factors, including transcription, translation, enzyme activity, and metabolite concentration, and revealed as reaction rates at steady state in metabolic pathways [20–22]. Therefore, modularization of metabolic network based on pattern of metabolic flux variation would cluster biochemical reactions, which can then be considered as functional units, and simplify subsequent computational analyses.
BN analysis is increasingly adopted to extract useful information from messy highthroughput biological data [23, 24]. It is a probabilistic graphical model that reveals changes that each variable causes in another variables, so called causal relationship, in the form of a directed acyclic graph, and has successfully been applied to reverse engineering of various biological networks from large datasets, including transcriptome and proteome data [25–27]. BN analysis, in this study, was used to predict causal relationships among reactions and subsequently their metabolic modules based on their metabolic fluxes.
With these tools, we herein conceptualized a f ramework for network m odularization and B ayesian network analysis (FMB) in order to investigate the effects of specific perturbation on metabolism and to explore other complementary features of the perturbed cell in the context of metabolism, which other currently available highthroughput techniques do not provide. This framework analyzes a biological system subjected to the specific perturbation by clustering reactions of similar functions (or metabolic flux variation pattern), and graphically displaying direction of influences among these clusters as a result of the specific perturbation. For this, FMB employs constraintsbased flux analysis, hierarchical clustering and BN analysis in sequence. For a demonstration, FMB was applied to the genetically perturbed Escherichia coli metabolism using its genomescale metabolic network model [28]. The final outcome is a BN that is a causal network that shows influential correlations among reactions and their metabolic modules that are critical to the metabolic adjustment in response to the specific perturbation.
Results and discussion
Step AC: Generation of many samples of genomescale metabolic flux data
Physiological constraints used in constraintsbased flux analysis for simulating metabolism of wildtype and lpdA knockout mutant of E. coli
Enzyme  Metabolism  Condition  v^{13C}or v^{ fmt } (mmol/g dry cell weight/h)  σ 

Phosphotransferase system for Dglucose transport  Transport  Wildtype  3.04  0.01824 
ΔlpdA mutant  2.48  0.1488  
Cell growth rate    Wildtype  0.20  0.0100 
ΔlpdA mutant  0.22  0.0110  
Glucose6phosphate isomerase  Glycolysis/Gluconeogenesis  Wildtype  2.39  0.3107 
ΔlpdA mutant  1.79  0.2327  
Pyruvate kinase  Glycolysis/Gluconeogenesis  Wildtype  1.09  0.0109 
ΔlpdA mutant  0.26  0.0026  
Glucose 6phosphate dehydrogenase  Pentose phosphate pathway  Wildtype  0.61  0.0732 
ΔlpdA mutant  0.64  0.0768  
Phosphogluconate dehydrogenase  Pentose phosphate pathway  Wildtype  0.61  0.0732 
ΔlpdA mutant  0.32  0.0384  
Phosphoenolpyruvate carboxylase  Anaplerotic reactions  Wildtype  0.67  0.0469 
ΔlpdA mutant  1.61  0.1127  
Phosphoenolpyruvate carboxykinase  Anaplerotic reactions  Wildtype  0.07  0.0070 
ΔlpdA mutant  0.93  0.0930  
Pyruvate dehydrogenase  Glycolysis/Gluconeogenesis  Wildtype  3.56  0.2136 
ΔlpdA mutant  0.00  0.0000  
αketoglutarate dehydrogenase  Citrate Cycle (TCA)  Wildtype  Not constrained  Not constrained 
ΔlpdA mutant  0.00  0.0000  
glycine cleavage system  Folate Metabolism  Wildtype  Not constrained  Not constrained 
ΔlpdA mutant  0.00  0.0000 
Subsequently, randomly generated errors were introduced to the ^{13}Cbased metabolic flux and cell culture data constraining the reactions, and constraintsbased flux analysis was iteratively performed with these randomly adjusted experimental constraints in order to generate a large number of data samples; as a result, constraintsbased flux analysis produces a different solution every time it is performed, satisfying the requirement that statistically reliable BNs demand many observations of the target system (see Methods). Infeasibility of optimization problem from applying too many constraints was handled with least absolute deviation method (see Methods). In this study, a total of 2,000 samples were generated at the end, each 1,000 sets of data from the wildtype (control condition) and its genetic mutant (perturbed condition), respectively. The resulting matrix, NbyM, is named flux matrix in this study, where N is the number of reactions and M is the number of samples; M is 2,000 in this study (Figure 1C). It is important to note that flux data samples generated for the two different conditions are combined into a single flux matrix in order to reflect the effects of perturbation on the metabolic flux distribution, which enables BN analysis to more reliably infer the directionality of causal relationships among reactions and their modules [26].
Step D: Selection of core reactions
In the flux matrix generated, the number of reactions needs to be reduced to the size that can be handled by BN analysis because only a fraction of reactions, namely metabolic core reactions, govern the metabolic activity [15], and the network model has many functionally duplicating reactions that show exactly the same metabolic flux values under all circumstances, and do not contribute to additional information. For this, a series of reduction processes were performed to select only the core reactions from the flux matrix (Figure 1D). The definition of core reaction in this study includes 1) a feature of significant flux variations accompanied by specific perturbation and 2) being almost always active under all samples of either control or perturbation condition [15]. Here, being almost always active refers to having nonzero fluxes for a reaction in more than 95% of samples under either control or perturbed condition, thereby becoming a core reaction. This shows a slight contrast with the previous definition that the core reactions be always active under all circumstances [15]. This new definition was to sufficiently cover the number of reactions suitable for BN analysis.
With this in mind, reactions were first selected, which were highly influenced by perturbation in terms of variation in flux values using a paired ttest; two sets of 1,000 flux values from control and perturbation condition were compared for each reaction. Reactions with P < 10^{2} were considered to be significantly influenced by the perturbation. Among significantly influenced reactions, reactions that have nonzero fluxes in more than 95% of samples under control or perturbed condition were selected, based on the aforementioned definition of core reaction, by which they coordinate and significantly contribute to the metabolic integrity [15]. Next, for duplicate reactions, reactions that have exactly the same flux values in all samples, only one of them was selected because they do not contribute to the robust statistical conclusion, but only increase the data volume (Additional file 1). The removed duplicate reactions were considered in the later stage of data analysis after all statistical calculations. The final remaining reactions, core reactions, are the major fluxes operating the cell metabolism, and significantly influenced by given perturbation (Additional file 1). Core reactions were then subjected to subsequent modularization and BN analysis.
Step E and F: Modularization of metabolic network based on flux variation pattern
Metabolism was then modularized by grouping core reactions with metabolically correlated functions using hierarchical clustering; this modularization enables us to look at modules of reactions as functional unities instead of individual reactions, thereby simplifying our subsequent analysis (Additional file 1). Here, flux variation was used as a criterion for clustering, such that reactions that have similar pattern of flux variation from sample to sample were grouped. This is analogous to clustering of genes based on their expression level in transcriptome data [29]. Specifically, the reduced flux matrix containing only the core reactions was converted to a fluxpattern matrix that represents the flux variation pattern from one sample to another (Figure 1EF) (see Methods). As a result, reactions in a cluster show synchronized patterns of increases (designated by the value ‘1’ in the fluxpattern matrix) and decreases (designated by ‘1’ in the fluxpattern matrix) in flux values in response to the perturbation, implying that they pursue correlative biological functions (see Methods for further details on generation of various fluxpattern matrices and their adjoined form). Be noted that the term cluster and module are interchangeably used to discuss the results henceforth.
Step G and H: Bayesian network analysis of metabolic modules
Finally, BN analysis was employed to predict causal relationships of reactions at local and global scale. At local scale, BN of reactions in a cluster was first inferred (Additional file 2), and representative reactions that most heavily influence other reactions in each metabolic cluster were identified by using the concept of mutual information (MI) [30] and total mutual information (TMI) [31, 32]; representative reactions have the highest value of TMI, and this procedure is detailed in Methods. Selected representative reactions from each module were then subjected to another round of BN analysis, yielding a global scale BN (Figure 1GH; Additional file 1). Because reactions in each module possess positively correlated functional features, or flux variation pattern, we speculate that BN with representative reactions would reveal a reasonable systemwide causal network that shows correlations among their corresponding metabolic modules as well in response to specific perturbations.
Although the FMB looks into correlations that exist in metabolism system, it is distinct from previously established relevant approaches, for instance flux coupling analysis [12], and uniform random sampling and subsequent calculation of correlation coefficients among reactions [14] in several aspects: (1) FMB reveals correlations among metabolic modules in addition to individual reactions by considering flux variation patterns of reactions using clustering, thereby showing modularlevel behaviours under specific condition; (2) causal relationships predicted from the FMB are graphically shown with arrows, which provides better readability; (3) FMB is more oriented to contextspecific biological issues, such as specific gene knockout, by imposing constraints of mutantspecific ^{13}Cbased metabolic flux and cell culture data. These distinctions of FMB would provide complementary aspects of the cell that the previously reported approaches do not provide.
Application of FMB to lpdA gene knockout mutant of E. coli
As a demonstration of the framework FMB, it was applied to the wildtype and lpdA mutant of E. coli to systematically evaluate how the genetic perturbation, lpdA gene knockout, affects E. coli metabolism, cultured in defined minimal media with glucose [33]. This lpdA gene encodes lipoamide dehydrogenase, which is an important component of pyruvate dehydrogenase complex, αketoglutarate dehydrogenase, and glycine cleavage system [34, 35]. Therefore, LpdA assumes an important biological role in metabolism at broad scope as a component of enzyme complexes: pyruvate dehydrogenase complex for connecting glycolysis and TCA cycle, and producing acetylCoA and CO_{2} under aerobic condition, αketoglutarate dehydrogenase for generating NADH from the operation of TCA cycle, and glycine cleavage system for one carbon metabolism that is associated with nucleotides and cofactors [33]. Consequently, such biological importance and global effects that LpdA is likely to exert on the cellular physiology made this enzyme component an ideal target for the application of FMB. Results of each step in FMB are represented as follows.
Step AC: Generation of flux matrix from metabolic network model
Flux matrix containing 979 reactions and 2,000 data samples were first generated from constraintsbased flux analysis with constraints of ^{13}Cbased metabolic flux and cell culture data under control and perturbation conditions, combined with random errors (Figure 1AC). It should be noted that flux values from lpdA knockout mutant were obtained by additionally constraining the flux values of pyruvate dehydrogenase complex, αketoglutarate dehydrogenase, and glycine cleavage system to zero in order to reflect the knockout of lpdA gene.
Step D: Selection of core reactions
Flux matrix was then subjected to a series of filtering processes; initial 979 reactions were reduced to 377 reactions from ttest, 213 reactions after filtering reactions that have nonzero fluxes in more than 95% of samples in either conditions, and finally 106 core reactions after removing duplicate reactions and selecting one of them for subsequent analysis (Figure 1D; Additional file 1).
Step E and F: Modularization of metabolic network from fluxpattern matrices
The reduced flux matrix was next converted to fluxpattern matrices (Figure 1E), and hierarchical clustering was performed (Figure 1F). A total of 44 clusters, or metabolic modules, were created with size of each cluster ranging from 1 to 30 reactions (Additional file 1 and 2). Duplicate reactions that were previously filtered in the flux matrix were also considered at this stage by inserting all duplicate reactions in the same cluster. Overall, reactions involved in the same submetabolism were grouped into the same cluster, partially demonstrating that flux variation pattern employed as a criterion for hierarchical clustering does well represent the functional aspect of reactions. Cluster 5 (cell envelope biosynthesis), cluster 11 (tyrosine, tryptophan, and phenylalanine metabolism), and cluster 15 (membrane lipid metabolism) are the exemplary clusters that have reactions from the same submetabolism. For clusters having reactions from different submetabolisms, biochemical link that connects these reactions together can be inferred. Cluster 28 belongs to this case; reactions from TCA cycle, such as succinate dehydrogenase, and those from oxidative phosphorylation, including cytochrome oxidase and ATP synthase, as well as oxygen transporter are clustered in the same module as they all are linked through aerobic electron transfer for respiration. Likewise, reactions from cell envelope biosynthesis and membrane lipid metabolism tend to be clustered together because fatty acids created from membrane lipid metabolism are fed into the biosynthesis of cell envelope. Cluster 42 is such a case, and has, in addition, arabinose5phosphate isomerase that produces arabinose5phosphate critical for lipopolysaccharide and alanine racemase that converts Lalanine to Dalanine, an essential building block of peptidoglycan. All these examples so far discussed indicate that reactions can be grouped together based on their functional characteristic that is flux variation pattern from sample to sample. Because each cluster represents functionally similar reactions, this modularization process will ease our subsequent interpretation of the data.
Step G and H: Generation of a global scale Bayesian network using representative reactions
Once the E. coli metabolism was modularized, representative reaction for each cluster was selected (Figure 1G), and BN analysis was applied to them in order to investigate the causal relationships existing among these clusters at global scale (Figure 1H; Additional file 1). Because each module is presented as a representative reaction in the global scale BN, other reactions in their corresponding cluster should also be considered during interpretation. Good starting point for the detailed interpretation would be nodes that contain reactions directly perturbed, genetically or environmentally, or those that have the highest number of links influencing others.
Analysis of the results from FMB
As cell growth rate is reduced upon knockout of lpdA [33], it is noteworthy to observe how the cluster 1 and 44, which contain reactions of cell growth rate and pyruvate dehydrogenase, respectively, are correlated in the global scale BN (Figure 2). Although these two clusters were both shown to influence cluster 29, and there is no direct influential path from cluster 44 to 1 or vice versa, we presume that reduction of the cell growth rate and fluxes of other functionally coordinated reactions in the cluster 1 were initially caused by the negatively affected reactions of cluster 29. This causal relationship becomes clearer when we consider the fact that phosphoglucomutase is the representative reaction of cluster 29, which converts glucose6phosphate to glucose1phosphate that is a precursor of glycogen and cell envelope, and highly associated with biomass formation. It was reported that excessive accumulation of glycogen in E. coli resulted in increased biomass formation, which, in some part, experimentally supports the correlation between cluster 1 and 29 [37]. In turn, cluster 29 seems to have been affected by cluster 44 as inferred in the global scale BN because lpdA knockout mutant cannot efficiently generate energy for biosynthesis of biomass components, compared to the wildtype, due to the blocked pyruvate dehydrogenase that hinders further oxidation of glucose for oxidative phosphorylation (Figure 3). Here, it should be noted that, along with glycogen, enzymes involved in rather wider range of metabolisms, including amino acids, cell envelope, and nucleotides in cluster 29, indicate that various spots in the metabolic network are likely to be influenced by the removal of lpdA, which, in turn, might have negatively affected the cell growth rate. As a result, the global scale BN reveals closer positioning of cluster 1 and 44 through cluster 29.
Finally, essentiality of reactions calculated with constraintsbased flux analysis provided additional complementary information on the clusters; essentiality has a scale of 0 to 1, and reactions with greater essentiality receives the value closer to 1 (Methods). First, reactions involved in clusters 1 to 44 had average essentiality of 0.744 and 0.748 for the wildtype and lpdA knockout mutant, respectively, while average essentiality of whole reactions was 0.215 and 0.216 under respective condition. This indicates that reactions collected into the clusters are heavily associated with cellular energy production, in comparison with those not included in the clusters. Closer examination of these clustered reactions reveals that clusters surrounding cluster 44 have relatively low essentiality (Figure 3), indicating that they are likely to be more committed to flux redistribution upon lpdA knockout than creating essential biomass constituents. This observation is consistent with the previous report, in that a set of reactions responsible for carbon flux distribution in E. coli are consistently included in cluster 44 and its surrounding clusters [38]. On the contrary, cluster 1 has higher number of surrounding clusters that have complete essentiality (= 1.000), which shows that they are more directly associated with biomass formation (Figure 4). Interestingly, although reactions in these 44 clusters cover most of the reactions in the metabolic core previously reported in Almaas et al., which are defined to be always active under all examined circumstances [15], most of those overlapping reactions were gathered in the clusters that are not under direct influence of the cluster 1 and 44. From this, it can be considered that FMB in this study identifies reactions that are under direct influence of the specific perturbation, lpdA knockout in this case, in addition to the reactions required for the survival of the organism, which corresponds to the metabolic core [15].
Taken together, FMB provides theoretical evidences for the possible correlations among metabolic modules in response to the perturbation, as exemplified by cluster 1, 29, and 44 in this study. This study would consequently reveal additional pictures of the cellular physiology that could complement conventionally available biological information.
Conclusions
FMB developed in this study groups functionally similar reactions, and subsequently investigates the causal patterns among metabolic clusters in response to specific perturbation. This newly generated knowledge of causal patterns enables us to capture significant changes in metabolism at the level of metabolic modules, which distinguishes itself from conventional MFA. Hence, it could be used as another tool for examining the microbial physiology in addition to currently available genomewide highthroughput techniques [39]. Furthermore, this FMB could be applied to interpreting different types of specific perturbations, including multiple gene knockouts and environmental stresses.
Methods
Overall scheme and rationale for each step of FMB are elaborated in Results and Discussion (Figure 1) while its critical computations steps that require extensive explanation are mentioned in this section. All the computations were performed using 2.80 GHz Intel i5 processors.
Genomescale metabolic network model of E. coli
In this study, a previously reported genomescale metabolic network model of E. coli was used [28]. This model is comprised of 979 reactions and 814 metabolites. Simulations were performed using the program package MetaFluxNet [40] and GAMS (GAMS Development Corp., Washington DC, USA) for the optimization technique developed in this study as detailed below.
Simulation of the metabolic network model with random generation of constraints based on ^{13}Cbased metabolic flux data
Simulation of the metabolic network model along with data preprocessing were implemented in GAMS for the wildtype (control condition) and lpdA mutant (perturbed condition), separately. For application of FMB to E. coli metabolism, ^{13}Cbased metabolic flux data were adopted from Li et al. [33] for wildtype (control) and lpdA mutant (perturbation) for genetic perturbation. All constraints were from the continuous culture of E. coli at dilution rate of 0.2 h^{1} under the defined minimal medium with glucose (Table 1). Here, instead of considering all the reactions, only those heavily responsible for the split of fluxes between major metabolic pathways were constrained with the ^{13}Cbased metabolic flux data (Table 1); they include glucose6phosphate isomerase, pyruvate kinase, glucose 6phosphate dehydrogenase, phosphogluconate dehydrogenase, PEP carboxylase, PEP carboxykinase and pyruvate dehydrogenase [38]. In addition, phosphotransferase system for glucose uptake and cell growth rate were also given constraints [33]. Metabolic fluxes of these constrained reactions were allowed to be reconciled within the range of their standard deviation (Table 1). In addition to pyruvate dehydrogenase above, flux values of αketoglutarate dehydrogenase and glycine cleavage system were also additionally constrained to zero for calculating metabolic flux data for lpdA gene knockout mutant.
where S is a stoichiometric matrix that consists of metabolite i in reaction j (i x j), and v is a vector showing the flux of reaction j (j x 1). α and β are lower and upper bounds for the flux of reaction j. K, a subset of J, is a set of reactions constrained with ^{13}Cbased metabolic flux data, and B, the other subset of J, is a set of reactions constrained with the rates of cell growth and metabolite secretion from the continuous cell culture. v^{ con } is a vector of fluxes calculated from constraintsbased flux analysis whose experimental values are available. v^{ 13C } is a vector of intracellular fluxes obtained from ^{13}Cbased MFA, and v^{ fmt } is a vector of the cell growth rate and glucose uptake rate from the continuous cell culture. v_{ k } and v_{ b } are constrained to be randomized values, being equivalent to v^{ 13C } and v^{ fmt } with randomly added errors, respectively. Finally, random(μ,σ) is a function that generates a random number according to normal distribution with mean μ and standard deviation σ. In this study, mean was set to the flux value from ^{13}Cbased MFA, v^{ 13C }, or the rates of cell growth and glucose uptake, v^{ fmt }, and standard deviation was differently set for each reaction according to the measurement error reported in the literature [43]. Above optimization formulation was iterated 1,000 times using linear programming for each wildtype (control condition) and lpdA mutant (perturbed condition), yielding a total of 2,000 sets of data. Flux data generated for both conditions were merged into a flux matrix, so that the directionality of causal relationships can properly be inferred by BN analysis [26].
Modularization of metabolic network based on flux variation pattern
For conversion of reduced flux matrix having only core reactions into fluxpattern matrix that contains information on the flux variation patterns, flux value of a reaction sample (m^{ th } column) in the reduced flux matrix was compared to that of its following sample (m+1^{ th } column), and ‘1’ is given in the position (n, m) of fluxpattern matrix if the flux value of the reaction in (n, m) is increased in (n, m+1) of the flux matrix. ‘1’ is given if the flux value is decreased in the following sample, and ‘0’ is given for the same flux values in the m^{ th } and m+1^{ th } sample (Figure 1E). This was done for all the reactions in the reduced flux matrix, completing a fluxpattern matrix. Therefore, the fluxpattern matrix uses ‘1’, ‘0’ or ‘1’ to represent the flux variation pattern of a reaction from sample to its following sample. Here, the fluxpattern matrix may be different, depending on the order of columns (samples) in the reduced flux matrix. Therefore, columns of the reduced flux matrix were randomly permutated, and the fluxpattern matrix was obtained accordingly (Figure 1E). Then, fluxpattern matrices independently generated from flux matrices with different order of samples (columns) were adjoined and subjected to hierarchical clustering. In this study, we tested adjoined matrices with different number of constituting individual fluxpattern matrices for hierarchical clustering in order to confirm that they produce consistent clusters of metabolic reactions. Adjoined 300 fluxpattern matrices showed agreeable consistency every time they were produced from metabolic flux data. Finally, it should be noted that fluxpattern matrices consisting of ‘1’, ‘0’ and ‘1’ were used only for clustering process. Once metabolic clusters were formed, previously calculated metabolic flux values (step A to C of FMB) corresponding to each clustered reaction were used in the subsequent BN analysis.
Conversion of reduced flux matrices with permutated samples (columns) into fluxpattern matrices, adjoining fluxpattern matrices and hierarchical clustering were conducted in Matlab (Mathworks). For hierarchical clustering, the “clusterdata” command was used. Euclidean distance, average linkage and cutoff value of 1.0 were used as calculation options.
Bayesian network analysis
One of the goals of this study is to infer such BNs from metabolic flux profiles. All the BN analyses were performed with commercially available software BayesiaLab (BAYESIA, Laval, France) [44, 45]. Firstly, a conditional probability table of all the nodes needs to be defined in order to infer BNs. Therefore, flux values need to be discretized into discrete variables in terms of probability. In BayesiaLab, equal frequency method was employed for this purpose, which preprocessed the flux values in 2,000 sets of metabolic flux data from control and perturbed conditions. This method divides the range of continuous attribute into intervals (four intervals in this study) where each interval contains equal or similar number of data points (i.e. flux values). Structure learning of BNs was then performed with tabu order algorithm for generating various candidate networks along with minimum description length (MDL) as a scoring criterion [46]. Structure learning is finding the network that best describes the causal relationships in the data, evaluated by specific scoring criterion. Tabu order algorithm searches the solution space of all possible candidate networks by adding, deleting or reversing arcs based on its specific rules. MDL then evaluates each BN with respect to the data, and selects the most likely one that describes the data. Here, MDL accounts for the tradeoff between accuracy and complexity of the learnt BN, and thus the optimal BN is both simple and accurate model of the data [46].
Identification of representative reactions in each metabolic module
t is the target node and r_{1,…,n} are the remaining nodes in the BN. This procedure was repeated by selecting other remaining reactions as the target node and calculating the TMIs likewise. At the end, a reaction with the highest TMI value was identified as the representative reaction. Representative reactions determined from each local scale BN were then subsequently subjected to another BN analysis altogether to infer causal relationships among metabolic fluxes on a global scale. From this procedure, global scale BN is completed, consisting of representative reactions for each metabolic module, thereby explicitly revealing possible influences existing among them.
Calculation of reaction essentiality
Essentiality for each reaction was obtained by calculating the cell growth rate with that reaction removed using constraintsbased flux analysis. The calculated cell growth rate for each reaction was then scaled (= 1 – g∆/g), wherein g∆ and g are the calculated cell growth rates of the wildtype and gene knockout mutant, respectively [47]. Essentiality of zero indicates that the removed reaction has no effects on the cell growth rate, while essentiality of one means complete stoppage of the cell growth rate for the removed reaction.
Notes
List of abbreviations
 BN:

Bayesian network
 FMB:

framework for network modularization and Bayesian network analysis
 MDL:

minimum description length
 MFA:

metabolic flux analysis
 MI:

mutual information
 PEP:

phosphoenolpyruvate
 TMI:

total mutual information.
Declarations
Acknowledgements
We thank Hongseok Yun for his thoughtful advises on clustering. This work was supported by the Advanced Biomass R&D Center of Korea (ABC20100029799) through the Global Frontier Research Program of the Ministry of Education, Science and Technology (MEST). Further support by World Class University program (R322008000101420) of MEST is appreciated.
This article has been published as part of BMC Systems Biology Volume 5 Supplement 2, 2011: 22nd International Conference on Genome Informatics: Systems Biology. The full contents of the supplement are available online at http://www.biomedcentral.com/17520509/5?issue=S2.
Authors’ Affiliations
References
 Oberhardt MA, Palsson BO, Papin JA: Applications of genomescale metabolic reconstructions. Mol Syst Biol. 2009, 5: 320PubMed CentralView ArticlePubMedGoogle Scholar
 Lee KH, Park JH, Kim TY, Kim HU, Lee SY: Systems metabolic engineering of Escherichia coli for Lthreonine production. Mol Syst Biol. 2007, 3: 149PubMed CentralView ArticlePubMedGoogle Scholar
 Alper H, Jin YS, Moxley JF, Stephanopoulos G: Identifying gene targets for the metabolic engineering of lycopene biosynthesis in Escherichia coli. Metab Eng. 2005, 7: 155164. 10.1016/j.ymben.2004.12.003.View ArticlePubMedGoogle Scholar
 Shen Y, Liu J, Estiu G, Isin B, Ahn YY, Lee DS, Barabasi AL, Kapatral V, Wiest O, Oltvai ZN: Blueprint for antimicrobial hit discovery targeting metabolic networks. Proc Natl Acad Sci U S A. 2010, 107: 10821087. 10.1073/pnas.0909181107.PubMed CentralView ArticlePubMedGoogle Scholar
 Folger O, Jerby L, Frezza C, Gottlieb E, Ruppin E, Shlomi T: Predicting selective drug targets in cancer through metabolic networks. Mol Syst Biol. 2011, 7: 501PubMed CentralView ArticlePubMedGoogle Scholar
 Kim HU, Kim SY, Jeong H, Kim TY, Kim JJ, Choy HE, Yi KY, Rhee JH, Lee SY: Integrative genomescale metabolic analysis of Vibrio vulnificus for drug targeting and discovery. Mol Syst Biol. 2011, 7: 460PubMed CentralView ArticlePubMedGoogle Scholar
 Kim HU, Kim TY, Lee SY: Genomescale metabolic network analysis and drug targeting of multidrug resistant pathogen Acinetobacter baumannii AYE. Mol Biosyst. 2010, 6: 339348. 10.1039/b916446d.View ArticlePubMedGoogle Scholar
 Thiele I, Palsson BO: A protocol for generating a highquality genomescale metabolic reconstruction. Nat Protoc. 2010, 5: 93121.PubMed CentralView ArticlePubMedGoogle Scholar
 Kim HU, Kim TY, Lee SY: Metabolic flux analysis and metabolic engineering of microorganisms. Mol Biosyst. 2008, 4: 113120. 10.1039/b712395g.View ArticlePubMedGoogle Scholar
 Park JM, Kim TY, Lee SY: Constraintsbased genomescale metabolic simulation for systems metabolic engineering. Biotechnol Adv. 2009, 27: 979988. 10.1016/j.biotechadv.2009.05.019.View ArticlePubMedGoogle Scholar
 Schellenberger J, Palsson BO: Use of randomized sampling for analysis of metabolic networks. J Biol Chem. 2009, 284: 54575461.View ArticlePubMedGoogle Scholar
 Burgard AP, Nikolaev EV, Schilling CH, Maranas CD: Flux coupling analysis of genomescale metabolic network reconstructions. Genome Res. 2004, 14: 301312. 10.1101/gr.1926504.PubMed CentralView ArticlePubMedGoogle Scholar
 Reed JL, Palsson BO: Genomescale in silico models of E. coli have multiple equivalent phenotypic states: assessment of correlated reaction subsets that comprise network states. Genome Res. 2004, 14: 17971805. 10.1101/gr.2546004.PubMed CentralView ArticlePubMedGoogle Scholar
 Price ND, Schellenberger J, Palsson BO: Uniform sampling of steadystate flux spaces: means to design experiments and to interpret enzymopathies. Biophys J. 2004, 87: 21722186. 10.1529/biophysj.104.043000.PubMed CentralView ArticlePubMedGoogle Scholar
 Almaas E, Oltvai ZN, Barabasi AL: The activity reaction core and plasticity of metabolic networks. PLoS Comput Biol. 2005, 1: e6810.1371/journal.pcbi.0010068.PubMed CentralView ArticlePubMedGoogle Scholar
 Barabasi AL, Oltvai ZN: Network biology: understanding the cell's functional organization. Nat Rev Genet. 2004, 5: 101113. 10.1038/nrg1272.View ArticlePubMedGoogle Scholar
 Hartwell LH, Hopfield JJ, Leibler S, Murray AW: From molecular to modular cell biology. Nature. 1999, 402: C4752. 10.1038/35011540.View ArticlePubMedGoogle Scholar
 Ravasz E, Somera AL, Mongru DA, Oltvai ZN, Barabasi AL: Hierarchical organization of modularity in metabolic networks. Science. 2002, 297: 15511555. 10.1126/science.1073374.View ArticlePubMedGoogle Scholar
 Ma HW, Zhao XM, Yuan YJ, Zeng AP: Decomposition of metabolic network into functional modules based on the global connectivity structure of reaction graph. Bioinformatics. 2004, 20: 18701876. 10.1093/bioinformatics/bth167.View ArticlePubMedGoogle Scholar
 Sauer U: Metabolic networks in motion: 13Cbased flux analysis. Mol Syst Biol. 2006, 2: 62PubMed CentralView ArticlePubMedGoogle Scholar
 Stephanopoulos G: Metabolic fluxes and metabolic engineering. Metab Eng. 1999, 1: 111. 10.1006/mben.1998.0101.View ArticlePubMedGoogle Scholar
 Nielsen J: It is all about metabolic fluxes. J Bacteriol. 2003, 185: 70317035. 10.1128/JB.185.24.70317035.2003.PubMed CentralView ArticlePubMedGoogle Scholar
 Needham CJ, Bradford JR, Bulpitt AJ, Westhead DR: Inference in Bayesian networks. Nat Biotechnol. 2006, 24: 5153. 10.1038/nbt010651.View ArticlePubMedGoogle Scholar
 Neapolitan RE: Learning Bayesian Networks. 2003, Harlow: Prentice HallGoogle Scholar
 Woolf PJ, Prudhomme W, Daheron L, Daley GQ, Lauffenburger DA: Bayesian analysis of signaling networks governing embryonic stem cell fate decisions. Bioinformatics. 2005, 21: 741753. 10.1093/bioinformatics/bti056.View ArticlePubMedGoogle Scholar
 Sachs K, Perez O, Pe'er D, Lauffenburger DA, Nolan GP: Causal proteinsignaling networks derived from multiparameter singlecell data. Science. 2005, 308: 523529. 10.1126/science.1105809.View ArticlePubMedGoogle Scholar
 Li Z, Chan C: Inferring pathways and networks with a Bayesian framework. FASEB J. 2004, 18: 746748.View ArticlePubMedGoogle Scholar
 Reed JL, Vo TD, Schilling CH, Palsson BO: An expanded genomescale model of Escherichia coli K12 (iJR904 GSM/GPR). Genome Biol. 2003, 4: R5410.1186/gb200349r54.PubMed CentralView ArticlePubMedGoogle Scholar
 D'Haeseleer P: How does gene expression clustering work?. Nat Biotechnol. 2005, 23: 14991501. 10.1038/nbt12051499.View ArticlePubMedGoogle Scholar
 Cover TM, Thomas JA: Elements of Information Theory. 1991, New York: WileyInterscienceView ArticleGoogle Scholar
 Nicholson AE, Jitnah N: Using mutual information to determine relevance in Bayesian networks. Proceedings of the 5th Pacific Rim International Conference on Artificial Intelligence: Topics in Artificial Intelligence: 22  27 November 1998; Singapore. Edited by: Lee HY, Motoda H. 1998, SpringerVerlag, 399: 410.
 Gibbons FD, Roth FP: Judging the quality of gene expressionbased clustering methods using gene annotation. Genome Res. 2002, 12: 15741581. 10.1101/gr.397002.PubMed CentralView ArticlePubMedGoogle Scholar
 Li M, Ho PY, Yao S, Shimizu K: Effect of lpdA gene knockout on the metabolism in Escherichia coli based on enzyme activities, intracellular metabolite concentrations and metabolic flux analysis by ^{13}Clabeling experiments. J Biotechnol. 2006, 122: 254266. 10.1016/j.jbiotec.2005.09.016.View ArticlePubMedGoogle Scholar
 Steiert PS, Stauffer LT, Stauffer GV: The lpd gene product functions as the L protein in the Escherichia coli glycine cleavage enzyme system. J Bacteriol. 1990, 172: 61426144.PubMed CentralPubMedGoogle Scholar
 Quail MA, Haydon DJ, Guest JR: The pdhRaceEFlpd operon of Escherichia coli expresses the pyruvate dehydrogenase complex. Mol Microbiol. 1994, 12: 95104. 10.1111/j.13652958.1994.tb00998.x.View ArticlePubMedGoogle Scholar
 Neidhardt FC, Umbarger HE: Chemical composition of Escherichia coli. Escherichia Coli and Salmonella: Cellular and Molecular Biology. Edited by: Neidhardt FC, Curtiss R. 1996, Washington, D.C.: ASM Press, 1316. 2Google Scholar
 Dedhia NN, Hottiger T, Bailey JE: Overproduction of glycogen in Escherichia coli blocked in the acetate pathway improves cell growth. Biotechnol Bioeng. 1994, 44: 132139. 10.1002/bit.260440119.View ArticlePubMedGoogle Scholar
 Sauer U, Eikmanns BJ: The PEPpyruvateoxaloacetate node as the switch point for carbon flux distribution in bacteria. FEMS Microbiol Rev. 2005, 29: 765794. 10.1016/j.femsre.2004.11.002.View ArticlePubMedGoogle Scholar
 Park JH, Lee SY, Kim TY, Kim HU: Application of systems biology for bioprocess development. Trends Biotechnol. 2008, 26: 404412. 10.1016/j.tibtech.2008.05.001.View ArticlePubMedGoogle Scholar
 Lee DY, Yun H, Park S, Lee SY: MetaFluxNet: the management of metabolic reaction information and quantitative metabolic flux analysis. Bioinformatics. 2003, 19: 21442146. 10.1093/bioinformatics/btg271.View ArticlePubMedGoogle Scholar
 Kim TY, Lee SY: Accurate metabolic flux analysis through data reconciliation of isotope balancebased data. J Microbiol Biotechnol. 2006, 16: 11391143.Google Scholar
 Bloomfield P, Steiger WL: Least Absolute Deviations: Theory, Applications, and Algorithms. 1983, Boston: BirkhauserGoogle Scholar
 Ishii N, Nakahigashi K, Baba T, Robert M, Soga T, Kanai A, Hirasawa T, Naba M, Hirai K, Hoque A: Multiple highthroughput analyses monitor the response of E. coli to perturbations. Science. 2007, 316: 593597. 10.1126/science.1132067.View ArticlePubMedGoogle Scholar
 Jouffe L, Munteanu P: New search strategies tailored for learning Bayesian networks. Proceedings of Tenth International Symposium on Applied Stochastic Models and Data Analysis: 12  15 June 2001; Compiegne. Edited by: Govaert G, Janssen J, Limnios N. 2001, The Societe Francaise de Statistique, 591596.Google Scholar
 Munteanu P, Bendou M: The EQ Framework for Learning Equivalence Classes of Bayesian Networks. Proceedings of the 2001 IEEE International Conference on Data Mining: 29 November  2 December 2001; San Jose. Edited by: Cercone N, Lin T, Wu X. 2001, IEEE Computer Society, 417424.View ArticleGoogle Scholar
 Lam W, Bacchus F: Learning Bayesian belief networks; An approach based on the MDL principle. Comput Intell. 1994, 10: 269293. 10.1111/j.14678640.1994.tb00166.x.View ArticleGoogle Scholar
 Edwards JS, Palsson BO: Robustness analysis of the Escherichia coli metabolic network. Biotechnol Prog. 2000, 16: 927939. 10.1021/bp0000712.View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.