Functional Annotation and Phenotype Characterization by Integrative Network Analysis

The rapid accumulation of biological network data translates into an urgent need for computational methods for graph pattern mining. In this talk, I will talk about several graph-based data mining methods which we developed to extract functional information from many massive biological networks. Here we focus on co-expression networks derived from microarray data. In the first project, we develop an efficient algorithm to identify recurrent patterns across multiple networks to discover biological modules, based on which we perform large-scale gene functional prediction. We further extend the algorithm to perform network biclustering, i.e. identifying condition-specific activation of network modules. In the second project, we design a graph mining approach to identify disease-specific network modules. Given two classes of graphs, one represents a particular disease (e.g. cancer), and another represents control. We characterize network modules which are frequently occurring in disease graphs versus control. Compared to commonly used microarray analysis approaches to identify differentially expressed genes between two phenotype classes, here we discover differentially regulated genetic networks. This addresses the important problem that phenotypic differences are often caused by interactions of genes, rather than individual genes.