Genomic best linear unbiased prediction using differential evolution

Abstract

In this paper we proposed a method to improve the accuracy of prediction of genomic best linear unbiased prediction (GBLUP). In GBLUP a genomic relationship matrix (GRM) is used to define the variance-covariance relationship between individuals and is calculated from all available genotyped markers. Instead of using all markers to build the GRM, which is then used for trait prediction, we used an evolutionary algorithm (differential evolution - DE) to subset the marker set and identify the markers that best capture the variance-covariance structure between individuals for specific traits. This subset of markers was then used to build a trait relationship matrix (TRM) that replaces the GRM in GBLUP (herein referred to as TBLUP). The predictive ability of TBLUP was compared against GBLUP and a Bayesian method (Bayesian LASSO) using simulated and real data. We found that TBLUP has better predictive ability than GBLUP and Bayesian LASSO in almost all scenarios.