Evaluation of Bagging approach versus GBLUP and Bayesian LASSO in genomic prediction

Abstract

The present study aimed to evaluate the predictive performance of bootstrap aggregating sampling technique (Bagging) in the context of genomic best linear unbiased prediction (GBLUP) method versus GBLUP and Bayesian least absolute shrinkage and selection operator (LASSO), in genomic prediction of livestock populations in different genetic architectures. For this purpose, different combinations of heritability (0.1 and 0.5), number of quantitative trait loci (QTL) (100 and 500) and distribution of QTL effects (normal, gamma, beta, Weibull and uniform) were considered. Also, a genome containing six chromosomes, 1 Morgan each, was simulated along which 1500 single-nucleotide polymorphism markers were evenly distributed. The prediction accuracies of the statistical models were obtained using the correlations between true (simulated) and predicted genomic breeding values. Results showed that, in different scenarios, the prediction accuracy using the GBLUP method was higher than that of the Bagging method (P > 0.05). When the heritability of the trait, the number of QTL and the distribution of QTL effects were 0.1, 500 and gamma, respectively, the prediction accuracy of Bagging and GBLUP indicated the highest similarity (P = 0.995). With low heritability (0.1) and low number of QTL (100), the maximum superiority of the Bagging method compared to the Bayesian LASSO method was obtained, which was statistically significant only when the distribution of QTL effects followed a gamma distribution (P < 0.05). For all three methods, the prediction accuracies decreased as the generation distance between the test and the reference generation increased (P < 0.001). In high heritability and when the QTL effects followed the Weibull distribution, all the three methods showed the highest prediction accuracy. In scenario of low heritability (0.1), low number of QTL (100) and gamma distribution for QTL effects, the difference between GBLUP and Bayesian LASSO methods as well as Bagging and Bayesian LASSO methods were statistically significant (P < 0.05. No significant differences were observed between the studied methods in other scenarios (P > 0.05). The results suggest that when the data are stable, the parametric (GBLUP and Bayesian LASSO) methods provide high prediction accuracy and it is not recommended to use the resampling (Bagging) method.