A Decomposition Approach for Discovering Network Building Blocks∗

The increasing availability of biological networks (proteinprotein interaction graphs, metabolic and transcriptional networks, etc.) is offering new opportunities to analyze their topological properties and possibly gain new insights in their design principles. Here we concentrate on the problem of de novo identification of the building modules of networks, which we refer to as network modules. We propose a novel graph decomposition algorithm based on the notion of edge betweenness that discovers network modules without assuming any a priori knowledge. We claim that the knowledge of the distribution of network modules carries more information than the distribution of subgraphs which is commonly-used in the literature. To demonstrate the effectiveness of the statistics based on network modules, we show that our method is capable of clustering more accurately networks known to have distinct topologies, and that the number of informative components in our feature vector is significantly higher. We also show that our approach is very robust to structural perturbations (i.e., edge rewiring) to the network. When we apply our algorithm to protein-protein interaction (PPI) networks, our decomposition method identifies highly connected network modules that occur significantly more frequently than those found in the corresponding random networks. Detailed inspection of the functions of the over-represented network modules in S. cerevisiae PPI network shows that the proteins involved in the modules either belong to the same cellular complex or share biological functions with high similarity. A comparative analysis of PPI networks against AS-level Internet graphs shows that in AS-level networks highly connected network modules are less frequent but more tightly connected with each other.