Comparison of parametric, semiparametric and nonparametric methods in genomic evaluation

Access to dense panels of molecular markers has facilitated genomic selection in animal breeding. The purpose of this study was to compare the nonparametric (random forest and support vector machine), semiparametric reproducing kernel Hilbert spaces (RKHS), and parametric methods (ridge regression and Bayes A) in prediction of genomic breeding values for traits with different genetic architecture. The predictive performance of different methods was compared in different combinations of distribution of QTL effects (normal and uniform), two levels of QTL numbers (50 and 200), three levels of heritability (0.1, 0.3 and 0.5), and two levels of training set individuals (1000 and 2000). To do this, a genome containing four chromosomes each 100-cM long was simulated on which 500, 1000 and 2000 evenly spaced single-nucleotide markers were distributed. With an increase in heritability and the number of markers, all the methods showed an increase in prediction accuracy (P<0.05). By increasing the number of QTLs from 50 to 200, we found a significant decrease in the prediction accuracy of breeding value in all methods (P<0.05). Also, with the increase in the number of training set individuals, the prediction accuracy increased significantly in all statistical methods (P<0.05). In all the various simulation scenarios, parametric methods showed higher prediction accuracy than semiparametric and nonparametric methods. This superior mean value of prediction accuracy for parametric methods was not statistically significant compared to the semiparametric method, but it was statistically significant compared to the nonparametric method. Bayes A had the highest accuracy of prediction among all the tested methods and, is therefore, recommended for genomic evaluation.