167
Genomic Prediction Accounting For Residual Heteroskedasticity
Classical genomic selection (GS) models that use single-nucleotide polymorphism (SNP) marker information to predict genetic merit of animals and plants usually assume homogeneous residual variance. However, this assumption seems questionable as environmental variability can be heterogeneous and it may affect the genetic control of a given quantitative trait. This study extends classical GS models, namely RR-GBLUP, BayesA, BayesB and BayesC, to explicitly account for residual heteroskedasticity using a hierarchical Bayesian mixed-models framework implemented with Markov Chain Monte Carlo methods. Competing GS models assuming homogeneous or heterogeneous residual variances were fitted to training data under simulation scenarios reflecting a gradient of increasing residual heteroskedasticity. Model fit of competing homoskedastic and heteroskedastic GS models was compared using prediction accuracy of genomic breeding values and pseudo-Bayes factors, both computed on a validation data subset one generation removed from the training dataset. Competing models were also fitted to two quantitative traits selected from a Michigan State University swine resource population dataset, namely carcass temperature and loin muscle pH 45 min after slaughter. These traits had been pre-screened for homoskedasticity and heteroskedasticity, respectively. Using a 5-fold cross-validation approach, competing GS models were compared based on predictive ability of phenotypes. Overall, under the conditions considered in this study, heteroskedastic GS models showed improved model fit and enhanced prediction accuracy compared to homoskedastic GS models under conditions of extreme residual variance heterogeneity; however, the magnitude of the improvement was too small (approximately 1% to 2% net gain in prediction accuracy) to confer practical relevance.
Keywords:
Genomic selection model, Heterogeneous residual variance, Genomic breeding values.