Application of Machine Learning Methods for Genome Wide Association Mapping in Agricultural Species Undergoing Strong Selection Pressure

Genome wide association studies (GWAs) are now widely used to estimate the association among single nucleotide polymorphisms (SNPs) distributed across the whole genome and one or more traits. This field has rapidly advanced through the availability of thousands of SNPs that can be genotyped in a cost effective manner.  Using appropriate methodology, and taking into account both the relationship among individuals and the correlation between the traits investigated, it is possible to enhance the power of GWAs studies, hence leading to more reliable results. Although genomic selection can decrease the intergeneration time and rapidly increase rates of genetic gain compared with traditional phenotypic selection, it is still important to pinpoint and localize the genetic variations that have a major effect on important traits and to test their interaction.  We have established a connection between statistical methods used in GWAs and machine learning approaches to take into account the effect, ranking and relationship among SNPs. Machine learning methods have been proposed  for  improving prediction and variable selection protocols in fields related to bioinformatics and computational biology, specifically through the use of the nonlinear and nonparametric Random Forest (RF) method. RF methodology can be used for both continuous and categorical traits and provides variable importance measures for variable selection purposes. Our results indicate that RF methods provide a reliable ranking of SNP effects that take indirectly in consideration interactions between linked and interacting SNPs, and is applicable for the discovery of genetic markers in agricultural species under strong selection pressure.