How robust are genomic selection methods?

Abstract

Genomic information from many single nucleotide polymorphism (SNP) markers can be used to increase the accuracy of estimated breeding values of young animals. This is termed genomic selection (GS) and is based on prediction of the effects of quantitative trait loci (QTL) in linkage disequilibrium (LD) with markers (Meuwissen et al. 2001). However, Habier et al. (2007) proposed that GS also relies on “relationships” between individuals to accurately predict genetic value. A better understanding of what GS actually predicts is needed to develop marker panels, training populations and methods for accurately estimating breeding values. The efficacy of methods used to predict genomic breeding value may depend on the underlying model of genetic variation, which is not well known. The aim of this study was to determine the accuracy and robustness of various methods used for genomic selection using a range of underlying genetic models and to compare the accuracy of GS when predicting one generation ahead (training set 1), several generations ahead (training set 2) or across different populations (training set 3). <br/> Three models of variation were used to simulate the genetic value of animals: (i) a QTL model in which few QTLs have a relatively large effect, (ii) a QTL model in which many QTLs have moderate effects and (iii) an infinitesimal model in which very many QTLs each have a very small effect. Genotype information from 60,000 markers was used to estimate the genetic value of animals using the following methods: (a) Bayes B, based on estimation of marker effects, (b) gBLUP, based on genomic relationships between animals and (c) traditional BLUP, based on pedigree relationships.