SNPs, Random Forests, and Asthma Susceptibility

• Activity Number: 371
• Type: Contributed
• Date/Time: Wednesday, August 6, 2003 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Alexandre Bureau*+ and Josee Dupuis and Kathleen Falls and Brooke Hayward and Tim Keith and Paul Van Eerdewegh
• Companies: Genome Therapeutics Corp. and Boston University School of Public Health and Genome Therapeutics Corp. and Genome Therapeutics Corp. and Genome Therapeutics Corp. and Genome Therapeutics Corp.
• Keywords: predictive importance ; genetic susceptibility ; classification trees

Abstract

A random forest (RF) is a collection of classification trees grown on bootstrap samples of observations, using a random subset of predictors to define the best split at each node. A class prediction for each observation is obtained by tallying the votes of the trees constructed without using the observation. The importance of a predictor is quantified by the increase in misclassification occurring when its values are randomly permuted among excluded observations. We extend this idea to pairs of predictors, to capture joint effects, and explore the properties of importance measures. The ability of RF to detect the equal importance of strongly correlated variables is attractive in the context of genetic association studies of very close single nucleotide polymorphisms (SNPs) that may show high correlation at the population level. We apply RF to the genotypes of asthma cases and unaffected controls at 42 SNPs in the asthma susceptibility gene ADAM33. Despite low prediction accuracy, a few SNPs have detectable importance. SNPs and SNP pairs highly associated with asthma tend to have the highest importance, but predictive importance and association do not perfectly correlate.