A Method for Learning Gene Association Networks from High-Dimensional Data

• Activity Number: 242
• Type: Contributed
• Date/Time: Tuesday, August 5, 2008 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Jie Cheng+ and Xiwu Lin and Kwan Lee*+
• Companies: GlaxoSmithKline and GlaxoSmithKline R&D and GlaxoSmithKline
• Keywords: partial correlation matrix estimation ; gene association network ; graphical Gaussian models ; Bayesian networks

Abstract

Inferring large scale gene networks from limited continuous data is a challenging problem in bioinformatics. This problem is closely related to partial correlation matrix estimation and graphical Gaussian model learning. We developed a java based tool for such task based on our previous work on learning Bayesian network from multinomial data. The basic idea is to learn the network structure in three phases: drafting, thickening and thinning. In drafting phase, marginal correlations are used to generate an initial guess of the structure. In thickening and thinning phases, low order partial correlations are used to constantly modify the network in a parsimony manner (in terms of both the number of tests and the order of the tests), guided by the network structure at that moment. Simulation data and public functional genomics data are used to evaluate the performance of the proposed method.