The effect of prior distributions of different statistical models on the accuracy of genomic prediction: simulation study

In the genomic selection, SNP markers across whole genome are used to estimate the marker effects. Genomic breeding value of animals can be predicted by different statistical methods. Genomic breeding value of animals can be predicted by different statistical methods. In the present study, accuracies of the predicted direct genomic breeding values were compared under several statistical models including Bayes A, B-LASSO gamma, B-LASSO beta and BGLR by considering two heritabilities of 0.3 and 0.05. Three values of π (0.1, 0.3, and 0.5) and four values of s2 (0.01, 0.1, 10, and 100) were simulated and correspondingly evaluated in terms of accuracy. The obtained results showed the highest accuracy of direct genomic breeding value was obtained under Bayes-A method, which were 0.88 and 0.69 for heritabilities of 0.3 and 0.05, respectively. Regression coefficients of methods for estimating the marker effects were more unbiased under heritability of 0.3 than 0.05. By considering the heritability of 0.3, the lowest and highest error were obtained under Bayes-A (121.2) and B-LASSO beta (165.2) methods, respectively. Under BGLR method, the highest and lowest accuracy of Bayes-A were obtained for π, 0.5 (0.81) and π, 0.1 (0.45). By increasing s2 parameter a decrease in the accuracy of genomic predictions was obtained. The obtained results suggested that to maximize the accuracy of the genetic prediction for traits with moderate to high heritability (0.3) optimal point of parameter π may ranged from 0.41 to 0.56 and while for traits with low heritability from 0.56 to 0.73.