A comparison of genomic selection methods for breeding value prediction

Recent advances in molecular genetics techniques have made dense marker maps available, and the prediction of breeding value at the genome level has been employed in genetics research. However, an increasingly large number of markers raise both statistical and computational issues in genomic selection (GS), and many methods have been developed for genomic prediction to address these problems, including ridge regression-best linear unbiased prediction (RR-BLUP), genomic best linear unbiased prediction, BayesA, BayesB, BayesC pi, and Bayesian LASSO. In this paper, these methods were compared regarding inference under different conditions, using real data from a wheat data set and simulated scenarios with a small number of quantitative trait loci (QTL) (20), a moderate number of QTL (60, 180) and an extreme number of QTL (540). This study showed that the genetic architecture of a trait should be fully considered when a GS method is chosen. If a small amount of loci had a large effect on a trait, great differences were found between the predictive ability of various methods and BayesC pi was recommended. Although there was almost no significant difference between the predictive ability of BayesC pi and BayesB, BayesC pi is more feasible than BayesB for real data analysis. If a trait was controlled by a moderate number of genes, the absolute differences between the various methods were small, but BayesA was also found to be the most accurate method. Furthermore, BayesA was widely adaptable and could perform well with different numbers of QTL. If a trait was controlled by an extreme number of minor genes, almost no significant differences were detected between the predictive ability of various methods, but RR-BLUP slightly outperformed the others in both simulated scenarios and real data analysis, thus demonstrating its robustness and indicating that it was quite effective in this case.