Performance evaluation of Boosting and Bayes A methods by different challenges of genomic architectures in discrete and continue traits

Genomic selection is a promising challenge for discovering genetic variants influencing quantitative and threshold traits for improving the genetic gain and accuracy of genomic prediction in animal breeding. In this study, performance of Boosting and Bayes A methods was investigated to evaluate genomic breeding values for binary threshold and quantitative traits in different marker densities using different genomic architectures.

Genomic data were simulated by QMSim software to reflect variations in heritability (h2 = 0.1 and 0.3), linkage disequilibrium (LD=low and high), number of QTL (QTL=150 and 450) and marker densities (10k and 50k) for 30 chromosomes. To create discrete threshold phenotypes in training set, individuals per generation were ranked ascending order according continuous phenotypes of QMSim output. Afterwards, depending on average simulated population, the threshold phenotype of individuals was define was code 0 (higher than average trait) and code 1 (lower than average trait). Eventually, genomic estimated breeding values were calculated using Bayes A and Boosting methods to evaluate accuracy of genomic prediction for threshold and continue traits.

Comparing to Bayes A method, Boosting algorithm was showed a wide range of genomic accuracy to changes marker density. Comparing to threshold Bayes A method, Boosting algorithm demonstrated an increase of 6.3 and 7.3 percentage on genomic accuracy of threshold traits when 10k and 50k SNPs panels were used, respectively. For traits with continue phenotypic distribution, performance of Bayes A was much more than Boosting, especially when the sparse panels were used. The structure of genomic architecture including heritability, number of QTL and LD were the most important factors affecting the accuracy of genomic prediction using Bayes A and Boosting methods. In this way, impact of heritability on performance of each of these models was more evident. Overall, genomic accuracies of Bayes A and Boosting methods showed more sensitive to QTL and LD fluctuations, respectively. For threshold traits with high density marker panels, the highest and lowest of genomic accuracy were obtained using Boosting (0.598) and Bayes A (0.510) methods, respectively, when the data set containing a lot of QTL was applied. For continue traits, the highest and lowest of genomic accuracy were obtained using Bayes A (0.702) and Boosting (0.569) methods, respectively, when the data set containing a few of QTL was used. the positive effect of increase LD on accuracies of genomic prediction of Boosting and Bayes A for the sparse panels was much more noticeable than high density panels.

The general trend of the present results indicated that Boosting and Bayes A methods showed their best performance for threshold and continue traits, respectively.