The Impact of Genetic Architecture on Genome-Wide Evaluation Methods

The rapid increase in high-throughput single-nucleotide polymorphism data has led to a great interest in applying genome-wide evaluation methods to identify an individual''s genetic merit. Genome-wide evaluation combines statistical methods with genomic data to predict genetic values for complex traits. Considerable uncertainty currently exists in determining which genome-wide evaluation method is the most appropriate. We hypothesize that genome-wide methods deal differently with the genetic architecture of quantitative traits and genomes. A genomic linear method (GBLUP), and a genomic nonlinear Bayesian variable selection method (BayesB) are compared using stochastic simulation across three effective population sizes and a wide range of numbers of quantitative trait loci (N_QTL). GBLUP had a constant accuracy, for a given heritability and sample size, regardless of N_QTL. BayesB had a higher accuracy than GBLUP when N_QTL was low, but this advantage diminished as N_QTL increased and when N_QTL became large, GBLUP slightly outperformed BayesB. In addition, deterministic equations are extended to predict the accuracy of both methods and to estimate the number of independent chromosome segments (M_e) and N_QTL. The predictions of accuracy and estimates of M_e and N_QTL were generally in good agreement with results from simulated data. We conclude that the relative accuracy of GBLUP and BayesB for a given number of records and heritability are highly dependent on M_e, which is a property of the target genome, as well as the architecture of the trait (N_QTL).THE rapid progress and reducing costs of genome sequencing and high-throughput DNA techniques have led to a great interest in applying genome-wide evaluation methods to identify individuals of high genetic merit. Genome-wide evaluation uses associations of a large number of SNP (single nucleotide polymorphism) markers across the whole genome with phenotypes to produce accurate estimates of breeding values (EBVs) for candidates to selection (Meuwissen et al. 2001). The accuracy of genome-wide selection (i.e., selection based on genomic EBVs) is expected to be substantially higher than that of traditional best linear unbiased prediction (BLUP) selection, which is based on pedigree and phenotypic data (Daetwyler et al. 2008; Goddard 2009; Hayes et al. 2009c). In addition, genome-wide selection has the potential to reduce inbreeding rates because of the increased emphasis on own rather than family information (Woolliams et al. 2002; Daetwyler et al. 2007; Dekkers 2007). Furthermore, the application of genome-wide evaluation approaches can significantly aid our understanding of quantitative trait genetic architecture.The genome-wide evaluation methods suggested to date can be broadly categorized into groups according to whether there is an assortment of the SNP by magnitude of effect or contribution to the variance. One group treats SNP homogeneously and includes variants of genomic best linear unbiased prediction (GBLUP). This group includes a form of ridge regression (Meuwissen et al. 2001) and the use of a realized relationship matrix computed from the markers instead of the traditional pedigree matrix (NejatiJavaremi et al. 1997; Villanueva et al. 2005; Hayes et al. 2009c). Both approaches have been shown to be equivalent (Habier et al. 2007; Goddard 2009). A second group provides for heterogeneity among SNP contributions to the variance, with some contributions permitted to be large while the remainder are small, possibly zero. This assortment is helped by Bayesian approaches, which place priors on numbers of SNP with major contributions (e.g., BayesA and BayesB; see Meuwissen et al. 2001, 2009; Lee et al. 2008), or with some penalty based on functions of the magnitude of effect for each SNP (e.g., Lasso; see Tibshirani 1996; Yi and Xu 2008) or with other smoothing metrics (Long et al. 2007). A third group attempts to reduce dimensionality by using principal components or partial least squares (Raadsma et al. 2008; Solberg et al. 2009) to identify an informative subset of SNP genotypes. The main two methods currently used in real data sets are a linear prediction method, GBLUP, and variants of nonlinear Bayesian variable selection approaches such as BayesB.In most simulated published data, the accuracy of BayesB outperformed that of GBLUP (e.g., Meuwissen et al. 2001; Habier et al. 2007; Lund et al. 2009). However, real data results have not consistently supported this conclusion. Two reviews of empirical results in dairy cattle to date have shown that GBLUP and BayesB result in very similar accuracies for most traits (Hayes et al. 2009a; Vanraden et al. 2009). One reason for the disagreement between simulated and real data results could be that the genetic architecture simulated is significantly different from what is found in real populations. Most studies published to date that compare methods using simulated architectures have considered only 50 or fewer QTL affecting the trait (e.g., Meuwissen et al. 2001; Habier et al. 2007; Lund et al. 2009). In this article we hypothesize that the relative utility of genome-wide evaluation methods depends significantly on both the genomic structure of the population and the genetic trait architecture.The main objective of this study was to compare a linear method, GBLUP, and a nonlinear variable selection method, BayesB, using simulated data across a range of population and trait genetic architectures to further understand the mechanics of genome-wide evaluation methods. An important secondary objective was to extend deterministic prediction models to predict the accuracy of both methods. Theoretical models complement stochastic simulation by helping the understanding of the factors involved in genome-wide evaluation performance and, in return, stochastic simulation is used to confirm theoretical derivations.