Using Random Forests to Identify Candidate Genes for Binary Trait in Poultry GWAS Datasets: The IMPACT of Correcting for Family Structures

Analyses of high-throughput genomic data present enormous challenges, due to the “large P, small N” and correlated population structure problems. One of the machine learning methods, Random Forests (RF), has been shown to be effective in addressing these problems. In this study, we examined the effectiveness of RF methodology in a genotype dataset derived from a commercial poultry population of about 1000 animals genotyped for over 50K SNPs. The trait of the interest is a binary trait - leg defect. The results from RF analysis were compared with a conventional GWAS method of logistic regression model. The power of RF was investigated at different parameter settings (size of forest trees and number of markers for each sampling) as well as with and without correcting for population structures. The results demonstrated the advantage of RF over the traditional GWAS approach, and the need for correcting for population structures prior to RF analysis.