Consistency analysis of metabolic correlation networks
© Müller-Linow et al. 2007
Received: 08 March 2007
Accepted: 24 September 2007
Published: 24 September 2007
Skip to main content
© Müller-Linow et al. 2007
Received: 08 March 2007
Accepted: 24 September 2007
Published: 24 September 2007
Metabolic correlation networks are derived from the covariance of metabolites in replicates of metabolomics experiments. They constitute an interesting intermediate between topology (i.e. the system's architecture defined by the set of reactions between metabolites) and dynamics (i.e. the metabolic concentrations observed as fluctuations around steady-state values in the metabolic network).
Here we analyze, how such a correlation network changes over time, and compare the relative positions of metabolites in the correlation networks with those in established metabolic networks derived from genome databases. We find that network similarity indeed decreases with an increasing time difference between these networks during a day/night course and, counter intuitively, that proximity of metabolites in the correlation network is no indicator of proximity of the metabolites in the metabolic network.
The organizing principles of correlation networks are distinct from those of metabolic reaction maps. Time courses of correlation networks may in the future prove an important data source for understanding these organizing principles.
In molecular biology, the consideration of biochemical processes as elements in an abstract network has become more and more important in the last few years [1, 2]. These network-oriented approaches bridge the gap between single units and collective behavior. Metabolism is, in a sense, the mediator between organisms and their environment. Resources, external conditions (represented by control parameters like temperature and concentrations of external agents) and energy all affect the organism via its metabolism. At the same time, metabolism is a key field of application of network biology. The classical approach considers metabolic networks as the pattern of connections of metabolites via enzyme-driven reactions. In this way, reaction networks are straightforward abstractions of what is commonly known as metabolic pathway maps. Global structural properties [3–6], statistical parameters like the degree distribution [3, 7] and the diameter [3, 8], and local properties like the motif content  are well studied. In combination with elementary flux mode analysis [10–14], possible routes between different metabolites are quantified within a metabolic map, while flux balance analysis (FBA) [15, 16] is suitable to predict the whole-cell behavior by adding constraints to the regulation of metabolic transformations. These theoretical approaches constitute important steps towards dynamics and have the potential to elucidate the fundamental link between topology and dynamics even further. Particularly the recent work by Almaas et al.  points in this direction.
Other types of metabolic networks have been established as well in recent time as the orthogonal networks, where enzymes are connected to each other when they share a common metabolite , and the correlation networks. In correlation networks a connection between two metabolites (nodes) represents an above-threshold correlation in metabolite concentrations.
Due to their quality of being derived from metabolic concentrations, they constitute an interesting intermediate between topology and dynamics. Here we study the compatibility of this intermediate with its two antipodes: the topological structure given by the network of metabolic reactions and the dynamic behavior given by the time evolution of the correlations between metabolites.
Such correlation networks have been reconstructed both from experimental data [20, 21] and from numerical simulations . Steuer et al.  used a stochastic system of linear equations based on an underlying metabolic network of biochemical reactions to discover correlations between metabolite fluctuations around a steady-state. This system is related to the metabolic control analysis (MCA) [10, 22–26], which also served as a groundwork for linear and non-linear perturbation studies of Camacho et al.  and Vance et al. . Camacho et al.  point out that different sources of variability can, in principle, lead to the observed correlations. All these studies show that the relation between metabolic correlation networks and the actual network given by the metabolic reactions is far from trivial. We therefore look at the similarity of these two graphs with simple topological tools asking, if two nodes are close to each other in the one network, when they are close in the other.
From the experimental point of view, metabolomic technologies provide widely-used tools to identify compounds in biological samples and to describe the current state of a system [29, 30]. For details on the quantitative and qualitative analysis of the single metabolites see Weckwerth and Morgenthal  and Kell . More than 1000 different compounds have been isolated and identified in a single tissue in this way.
The studies on different samples revealed strong correlations between certain pairs of metabolites, while most other combinations displayed little or no correlation. Such a correlation profile may serve as a basis for the construction of a metabolic correlation network. Because anti-correlations might have a physiological cause as well, they are treated equally to positive correlations. Beyond the identification of compounds via the concept of correlation networks, metabolomics is also capable to describe physiological processes in consequence of development and changing environmental conditions, since tissue samples can be analyzed at any point of time.
The interrelation between the architecture of a metabolic pathway map and the dynamic processes taking place upon it has sparsely been studied. Results concern the distribution of node degrees and of metabolic fluxes which have both been found to be scale-free [2, 3, 33] and the metabolic fluxes within the core of the metabolic network of E. coli, H. pylori, and S. cerevisiae , but most other studies deal with artificial networks. In the case of metabolic networks one observes discrepancies between theoretical graphs and the real-world networks' topologies. One key publication on gene expression data in yeast  indicates that metabolic reaction networks display both scale-free and exponential degree distributions depending on whether one takes all potential paths or particular realizations under certain conditions into account. Furthermore, the biochemical modules derived from experimental data deviate from proposed modules in theoretical reaction networks. Here we will not focus on topological properties of metabolic correlation networks, but look at them as a mediator between topology and dynamics. Since a correlation network represents a dynamic aspect of the metabolic network, it is suitable for comparison with its topological counterpart.
When analyzing the metabolite correlations in different plant samples at different sampling points in time one has to consider three potential contributions to the data: Short-term fluctuations in the metabolites' concentrations represented by all plant samples at a given time point may either reflect intrinsic noise or may originate from plant-to-plant variability. Lastly, systematic changes of the steady-state correlation networks over time along a diurnal cycle may contribute to the data. Here we want to find out, whether the (steady-state) correlation networks obtained from different time points are systematically similar. Two main effects can be expected: (1) network similarity should on average be higher for networks at neighboring time points compared to more distant time points, (2) day-night and night-day transitions should be associated with substantially lower network similarity (compared to neighboring time points at constant illumination).
Surprisingly, we find a high network similarity for the night-day transition and a low similarity for the day-night transition. Understanding this would, however, require more data. Omitting day-night transitions from this analysis and determining the consistency parameter θ at 0.6 <κ < 0.75 consequently yields a more pronounced result with comparable p-values.
On the basis of the data set of Ma and Zeng  and the data set of Weckwerth et al. [20, 21], 39 common compounds have been identified, which are present in both networks. This value primarily depends on the number of identified compounds in the metabolomic analysis. Unidentified compounds are systematic signals in the metabolomic data, which cannot be unambiguously associated with a node in the reaction network. The effect of unidentified metabolites is difficult to estimate. Whether they influence the length scale (an average distance of identified metabolites) or not mainly depends on the individual structure of the network. The known metabolites show approximately the same distribution of pair distances as the complete reaction network and can therefore be regarded as a representative subset of the whole graph in this property. We also checked the influence of unknown nodes on the network similarity σ for two Erdős-Rényi (ER) random graphs  (with a starting value of σ = 0.3) when using less and less nodes for this computation. Similarity σ displays only a marginal dependency on the number of contributing nodes even if half of the nodes are used.
Selected common compounds from the metabolic reaction and the metabolic correlation networks
In this work we investigated systematically the relationship between metabolic correlation networks and genome-wide predicted reaction networks. In recent studies we investigated how correlation networks are causally connected to the underlying biochemical reaction network and its regulation [19, 20, 30, 37]. However, all these studies revealed the discrepancy between predicted pathway connectivity and a simple extrapolation of these hypothetical networks to correlation networks. We detected overall changes in the correlation network topology, which are rather based on specific enzyme activity alterations [21, 30]. Therefore a systematic comparison of predicted pathway map structure and the experimental metabolite covariance is crucial. Based on the prediction that alterations in biochemical regulation will change the correlation network structure  we have investigated different correlation network structures during a diurnal time course in Arabidopsis plant leaf samples. Intriguingly, the structure of these networks did not match the predicted pathway connectivity in plant metabolism, but nevertheless varies systematically in time. Moreover, compared to theoretical pathways in Arabidopsis the correlation network structure revealed different and novel clusters of compounds such as the correlation of aromatic amino acids and housekeeping sugars which is not predicted by the topology of theoretical pathways (Fig. 6). This fact supports our hypothesis that the correlation network largely reflects biochemical regulation. Particularly this observation that the systematics of correlation networks are somewhat orthogonal to those contained in pathways networks, suggests that the systematic and differential analysis of metabolite correlation network structures and metabolite covariances will in the future lead to novel insights into biochemical regulation and regulatory hubs in the in vivo system.
The analysis of the correlation data and the respective correlation networks was based on the experimental data sets of Weckwerth et al. [20, 21]. In their studies, they recorded the concentrations of 257 different metabolites of Arabidopsis thaliana. Ten samples of this plant had been cultivated for 26 days under the same conditions with a diurnal rhythm of 8 h light and 16 h darkness. Data from 6 sampling points (three in each phase of illumination; in detail: 0.5 h, 4 h, and 7.5 h light; 0.5 h, 8 h, 15.5 h dark) were drawn from each plant, resulting in 60 measurements.
In principle, there are two practicable approaches to construct a correlation network from the data given here. Considering long-term fluctuations one would use the plant-specific changes in time given by six concentration values. We preferred a statistically more reliable method which considers the short-term fluctuations provided by 10 different plant samples. Using 6 samples in time instead of 10 plant samples largely limits the statistically reliable range of κ, setting the lower boundary of κ to a value of 0.82 (for a p-value of 0.05). This threshold matches the upper threshold approximately, where all networks tend to fragment.
By construction, the threshold (0 ≤ κ ≤ 1) regulates the connectivity of a correlation network. For κ = 0 one obtains a complete network, i.e. every node is connected to every other node. Occasionally, in the measurements some metabolites have not been registered. In these cases gaps occur in the data matrix. We account for these missing measurements by introducing a second threshold ω into our analysis and restrict the range of metabolites in the network to those pairs with at least ω measurement results. In this way we qualitatively vary the reliability of the data entering our analysis. We changed this parameter in an interval of 4 ≤ ω ≤ 8 in order to validate the robustness of the results with regard to the choice of data.
Various databases based on genomic analysis provide information on metabolic reactions. Among these the KEGG database  has been used by several research groups for the reconstruction of metabolic networks. This comparatively reliable group within the biochemical networks opens up a reaction space, which is used for metabolic transformations. Basically two approaches to interpret a reconstructed metabolic network exist so far, which differ in the consideration of ubiquitous compounds like H2O, ATP, and NADH, often referred to as current (or currency) metabolites. These hub forming compounds drastically reduce the average path length and, consequently, the pathways in such a metabolic network do not resemble the conventional order of reactions. Ma and Zeng  addressed this issue by eliminating all connections from a current metabolite, if the connections represent particular reactions, like the transfer of electrons or of certain functional groups. Additionally, they corrected mistakes and inconsistences in the original data set and checked the reversibility and direction information of each reaction.
The reconstruction of the metabolic network of A. thaliana, used in this paper, is based on the data from Ma and Zeng (http://genome.gbf.de/bioinformatics/index.html). This connection table displays 2070 different enzyme-driven transformations summed up over 107 organisms and broken down into reversible and irreversible reactions. A reaction in this table represents a potential connection within the metabolic network, while the existence of a link in a certain organism is indicated by a value greater than or equal to unity for each organism-reaction combination. (This value refers to the number of genes encoding this enzyme.) After cross-linking all metabolites listed in this table, an undirected subnetwork was extracted from this data by conversion of all directed links (irreversible reactions) into undirected links (reversible reactions) and by identification of the giant cluster, which is the largest connected component in a fractionated network. This step from directed to undirected networks for the sake of this comparison is necessary, as the correlation networks are undirected by construction. Thus, the 760 connections (438 reversible and 332 irreversible reactions) in the original data set were reduced to 492 undirected edges within a metabolic network of 404 nodes. This resulting network will be termed metabolic reaction network.
Various methods for the determination of the similarity of two networks exist so far. A prominent example is the graph alignment [39, 40], which is a suitable method for related networks. The great variability of the considered networks here asks for other statistical parameters. Thus, the determination of the similarity σ will be accomplished via comparison of all pair distances in the networks. The pair distance δ ij is defined as the shortest path between the two nodes i and j, i.e. the smallest number of links which connects two nodes. Undirected connected networks of size n hold a maximum number m max of pair distances of m max = n(n - 1)/2. This number drops in fractionated networks where the path lengths between nodes in different clusters cannot be calculated. For network comparison, only those pair distances are taken into account, which can be computed in both networks. The similarity σ = σ (G 1 , G 2) between the two networks G 1 and G 2 is the Pearson correlation coefficient of the respective pair distance vectors.
We tested this method both with real networks of different size but similar structure and with artificial graphs of the same size and connectivity which were altered systematically. Therefore, we compared the reaction network of A. thaliana with six reaction networks of other eukaryotes from the Ma and Zeng database. Despite their different size (332 <n < 625) the networks displayed an average network similarity σ of 0.81.
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.