Selection of important variables by statistical learning in genome-wide association analysis

Genetic analysis of complex diseases demands novel analytical methods to interpret data collected on thousands of variables by genome-wide association studies. The complexity of such analysis is multiplied when one has to consider interaction effects, be they among the genetic variations (G × G) or with environment risk factors (G × E). Several statistical learning methods seem quite promising in this context. Herein we consider applications of two such methods, random forest and Bayesian networks, to the simulated dataset for Genetic Analysis Workshop 16 Problem 3. Our evaluation study showed that an iterative search based on the random forest approach has the potential in selecting important variables, while Bayesian networks can capture some of the underlying causal relationships.