Discovering weighted motifs in gene co-expression networks

A important dimension of complex networks is embedded in the weights of its edges. Incorporating this source of information on the analysis of a network can greatly enhance our understanding of it. This is the case for gene co-expression networks, which encapsulate information about the strength of correlation between gene expression profiles. Classical unweighted gene co-expression networks use thresholding for defining connectivity, losing some of the information contained in the different connection strengths. In this paper, we propose a mining method capable of extracting information from weighted gene co-expression networks. We study groups of differently connected nodes and their importance as network motifs. We define a subgraph as a motif if the weights of edges inside the subgraph hold a significantly different distribution than what would be found in a random distribution. We use the Kolmogorov-Smirnov test to calculate the significance score of the subgraph, avoiding the time consuming generation of random networks to determine statistic significance. We apply our approach to gene co-expression networks related to three different types of cancer and also to two healthy datasets. The structure of the networks is compared using weighted motif profiles, and our results show that we are able to clearly distinguish the networks and separate them by type. We also compare the biological relevance of our weighted approach to a more classical binary motif profile, where edges are unweighted. We use shared Gene Ontology annotations on biological processes, cellular components and molecular functions. The results of gene enrichment analysis show that weighted motifs are biologically more significant than the binary motifs.