Abstract
Undirected gene coexpression networks obtained from experimental expression data coupled with efficient computational procedures are increasingly used to identify potentially relevant biological information (e.g., biomarkers) for a particular disease. However, coexpression networks built from experimental expression data are in general large highly connected networks with an elevated number of false-positive interactions (nodes and edges). In order to infer relevant information, the network must be properly filtered and its complexity reduced. Given the complexity and the multivariate nature of the information contained in the network, this requires the development and application of efficient feature selection algorithms to be able to exploit the topological characteristics of the network to identify relevant nodes and edges. This paper proposes an efficient multivariate filtering designed to analyze the topological properties of a coexpression network in order to identify potential relevant genes for a given disease. The algorithm has been tested on three datasets for three well known and studied diseases: acute myeloid leukemia, breast cancer, and diffuse large B-cell lymphoma. Results have been validated resorting to bibliographic data automatically mined using the ProteinQuest literature mining tool.