Genomic prediction of growth in pigs based on a model including additive and dominance effects

Abstract

Independent of whether prediction is based on pedigree or genomic information, the focus of animal breeders has been on additive genetic effects or 'breeding values'. However, when predicting phenotypes rather than breeding values of an animal, models that account for both additive and dominance effects might be more accurate. Our aim with this study was to compare the accuracy of predicting phenotypes using a model that accounts for only additive effects (MA) and a model that accounts for both additive and dominance effects simultaneously (MAD). Lifetime daily gain (DG) was evaluated in three pig populations (1424 Pietrain, 2023 Landrace, and 2157 Large White). Animals were genotyped using the Illumina SNP60K Beadchip and assigned to either a training data set to estimate the genetic parameters and SNP effects, or to a validation data set to assess the prediction accuracy. Models MA and MAD applied random regression on SNP genotypes and were implemented in the program Bayz. The additive heritability of DG across the three populations and the two models was very similar at approximately 0.26. The proportion of phenotypic variance explained by dominance effects ranged from 0.04 (Large White) to 0.11 (Pietrain), indicating that importance of dominance might be breed-specific. Prediction accuracies were higher when predicting phenotypes using total genetic values (sum of breeding values and dominance deviations) from the MAD model compared to using breeding values from both MA and MAD models. The highest increase in accuracy (from 0.195 to 0.222) was observed in the Pietrain, and the lowest in Large White (from 0.354 to 0.359). Predicting phenotypes using total genetic values instead of breeding values in purebred data improved prediction accuracy and reduced the bias of genomic predictions. Additional benefit of the method is expected when applied to predict crossbred phenotypes, where dominance levels are expected to be higher.

Impact of fitting dominance and additive effects on accuracy of genomic prediction of breeding values in layers

“…Thus, total genetic values were not better predictors of phenotypes for BayesC either. Contrary to our results, Lopes et al (2015) found that predicting phenotypes using total genetic values instead of breeding values in purebred pig populations improved prediction accuracy and reduced the bias of genomic predictions. Vitezica et al (2011) showed that GS models do not take past selection into consideration when only selected animals are genotyped and included in the analysis, which can result in a bias of genomic predictions and also reduced accuracy.…”

“…Misztal (1997) reported that accurate pedigree-based estimation of dominance variance requires at least 20 times as much data as required for estimation of additive variance. In pedigree-based models, which use expected degrees of relatedness between relatives, dominance variance may be confounded with environmental covariance of full sibs (common environment shared by full sibs) and with maternal effects (Lynch & Walsh 1998;Hill et al 2008), resulting in inflation of the dominance estimates (Misztal & Besbes 2000). The pedigree used in the current study consisted of full-sib families, but including an independent random effect of dam did not substantially change the estimates of dominance variance for most traits (results not shown).…”

“…Contrary to our results, Lopes et al . () found that predicting phenotypes using total genetic values instead of breeding values in purebred pig populations improved prediction accuracy and reduced the bias of genomic predictions.…”

“…The R package 'nadiv' (Wolak 2012) was used to construct the D matrix. The dominance genetic relationship (D gh ) between individuals g and h was computed as (Lynch & Walsh 1998):…”

“…Moreover, the D matrix used in this study is an approximation in the presence of inbreeding (see Materials and methods). With inbreeding and dominance, the covariance between inbred individuals is no longer a function of only additive and dominance variance (Lynch & Walsh 1998). Using both simulated and real data, Ovaskainen et al (2008) showed that for inbred populations, the approximations that are commonly used to compute pedigree-based dominance relationships (Equation 3) can produce substantially biased estimates in deep pedigrees, mostly overestimating dominance variance.…”