The use of genetic markers has become widespread for prediction of genetic merit in agricultural applications and is a beginning to show promise for estimating propensity to disease in human medicine. This process is known as genomic prediction and attempts to model the mapping between an organism's genotype and phenotype. In practice, this process presents a challenging problem. Sequencing and recording phenotypic traits are often expensive and time consuming. This leads to datasets often having many more features than samples. Common models for genomic prediction often fall victim to overfitting due to the curse of dimensionality. In this domain, only a fraction of the markers that are present significantly affect a particular trait. Models that fit to non-informative markers are in effect fitting to statistical noise, leading to a decrease in predictive performance. Therefore, feature selection is desirable to remove markers that do not appear to have a significant effect on the trait being predicted. The method presented here uses differential evolution based search for feature selection. This study will characterize differential evolution's efficacy in feature selection for genomic prediction and present several extensions to the base search algorithm in an attempt to apply domain knowledge to guide the search toward better solutions.
Read
