CAMND: Comparative analysis of metabolic network decomposition based on previous and two new criteria, a web based application

Abstract

Metabolic networks can model the behavior of metabolism in the cell. Since analyzing the whole metabolic networks is not easy, network modulation is an important issue to be investigated. Decomposing metabolic networks is a strategy to obtain better insight into metabolic functions. Additionally, decomposing these networks facilitates using computational methods, which are very slow when applied to the original genome-scale network. Several methods have been proposed for decomposing of the metabolic network. Therefore, it is necessary to evaluate these methods by suitable criteria. In this study, we introduce a web server package for decomposing of metabolic networks with 10 different methods, 9 datasets and the ability of computing 12 criteria, to evaluate and compare the results of different methods using ten previously defined and two new criteria which are based on Chebi ontology and Co-expression_of_Enzymes information. This package visualizes the obtained modules via “gephi” software. The ability of this package is that the user can examine whether two metabolites or reactions are in the same module or not. The functionality of the package can be easily extended to include new datasets and criteria. It also has the ability to compare the results of novel methods with the results of previously developed methods. The package is implemented in python and is available at http://eslahchilab.ir/softwares/dmn.

Introduction

Metabolism is the set of vital chemical reactions in the cell. Metabolic networks are used to model the behavior of the metabolism in the cell. They consist of two main elements: metabolites and reactions. Although there are many available methods for the metabolite networks analysis, applying them on the large-scale networks may be difficult and time-consuming. Network decomposition methods suggest that if subnetworks can be characterized and determined properly, researchers can deal with these smaller and simpler subnetworks instead of the complete large-scale networks. Network decomposition and analyzing the modules obtained from different methods offer a better route to understand the organization of metabolite networks. It also creates simple and helpful models of these complex systems. Different methods with various goals have been proposed for partitioning metabolic networks. Since networks are a collection of metabolites and reactions, some methods decompose networks based on metabolites, and some others decompose it based on reactions. We briefly review the main ideas of some state of the art decomposition methods. In 2002, Schuster et al. proposed a method to decompose the metabolic networks using “Hubs”, which can be found based on the scale-free property of such networks (Schuster et al., 2002). In 2003, Holme et al. presented a method with a similar idea that yields a hierarchical clustering of subnetworks using betweenness centrality (Holme et al., 2003). Guimera and Amaral introduced the “modularity” concept as a measure of the quality of network decompositions. They also proposed a method with a “simulated annealing” technique to maximize modularity in 2005 (Guimera and Amaral, 2005). In 2006, Newman constructed a “modularity matrix” for each network, and then he showed that modularity could be expressed by eigenvectors of this matrix. Using this idea, he proposed a recursive method for the decomposition of metabolic networks (Newman, 2006). In 2007, Poolman et al. proposed a method for grouping reactions in metabolic networks based on the correlation between reaction flux value in the steady-state, called “reaction correlation coefficient” (Poolman et al., 2007). Verwoerd introduced the “Netsplitter” method that uses a global connection degree based on random walks on the network to avoid excessive fragmentation in 2011 (Verwoerd, 2011). In 2011, Siradaran et al. introduced a novel method by redefining the modularity concept based on retroactive interactions (Sridharan et al., 2011). This method uses a reaction-centric representation of the network, in which reactions are vertices and metabolites are edges. In 2014, Müller and Bockmayr presented a “module-finding” method that could not decompose the whole network, i.e., it can find several reaction modules that may cover only a limited number of the reactions (Müller and Bockmayr, 2014). In 2015, Ding et al. proposed a new overlapping community detection algorithm called NDOCD. This method splits the network iteratively by removing links in derived link communities, which were identified by utilizing the node clustering technique (Ding et al., 2016). In 2015, Arne C. Reimersa presented the “Muller2” method based on k-modules. A k-module is a sub-network with low connectivity to the rest of the network. This method applied k-modules recursively, which results in a hierarchical decomposition of the metabolic network (Reimers, 2015).

In sum, there are various methods for metabolic network decomposition. Thus, there should be a systematic framework to analyze and compare the results of methods. In this study, we use some criteria that had been proposed by Rezvan and Eslahchi (2017) and introduce two new criteria, which are defined elaborately in the following, to have an appropriate assessment framework.