[
International Worm Meeting,
2009]
We have analyzed the structure of experimentally elucidated and predicted gene regulatory networks, obtained by various experimental methods including microarray experiments, using quantitative methods. These networks consist of nodes that represent genes, and nodes are connected when represented genes are related by some transcription factor, whether clarified or not. Actually known networks were compared with fictitious networks based on quantitative properties that measure network structure, and values to discriminate real and fictitious networks were extrapolated. Analyzed properties include local density, path length, centrality, and the flux. The last property, flux, is a new structural property which is analogous to information flow over the network connections. The flux value is defined for each connecion among nodes, and the flux density distribution function is computed. Then the classification of network is based on the discrepancy of distribution from the distribution calculated from actual biological networks. The properties measure globally the networks, and masks details of individual links. Given a predicted gene regulatory network, although the reliability of individual transcription relationships are difficult to be estimated, global evaluation is statistically possible. The proposed method uses solely the structural parameters of network, and no explicit biological knowledge is included. Explicit biological information denotes function of proteins, gene sequence data, species name, for instance. Methods that depends heavily on such information may score higher discrimination rate of plausible networks from nonbiological networks. However, biological information is still incomplete, and may result in partial use of information, which seems to be difficult. The proposed network evaluation criteria is useful to reject unreal or nonbiological networks, and can be used to judge the plausibility of networks generated by gene regulatory network prediction systems.