Comparison of parametric and resampling methods in genetic evaluation of quantitative traits with different genetic structure

The objective of this study was to compare three parametric (GBLUP, BayesB and RKHS) and two resampling (Bagging GBLUP and Random Forest) statistical methods in genomic prediction of traits with different genetic architecture. A genome consisting of three chromosomes, 1 Morgan each, was simulated on which 5000 SNPs and 50, 100 and 200 QTLs were distributed. The substitutions effects of QTLs were modeled with normal, gamma and uniform distributions with a level of heritability equal to 0.30. The predictive performance of statistical models was evaluated using the correlation between predicted and true breeding values as well as the regression of predicted values on true breeding values. In the target population, Random Forest resulted in overestimation of estimated regression coefficients while GBLUP, BayesB and RKHS led to an underestimation of regression coefficients of true breeding values on predicted breeding values. In exception of Bagging GBLUP, the performance of all statistical methods was the same in three gene effect distributions. However, the performance of GBLUP and BayesB was better than other statistical methods. A reason for this superiority could be the additive architecture of simulated traits. In conclusion, GBLUP and BayesB were superior over resampling methods in genomic predictions.