Effect of QTL Number and Distribution Effects on Some Statistical Methods Genomic Prediction of a Threshold Trait

Advances in high-throughput assays for genotyping single nucleotide polymorphisms (SNP) led to using these dense markers for predicting genomic breeding values, called genomic selection, and have been revolutionized both animal and crop breeding programs. The accuracy of genomic predictions is determined by the size of the reference population, extent of relationships between selection candidates and the reference population, linkage disequilibrium among markers and QTLs, the method used to estimate marker effects, and genetic architecture of the trait. Most of stuides on this topic focus on continuous traits, but some of economical important traits in livestock are threshold. Therefore, a stochastic simulation was used to compare 125 different scenarios for a threshold trait based on five 10, 50, 150, 500 and 2000 underlying QTL numbers, five normal, uniform, t, gamma and Laplace distributions for QTL effects and five BayesA, BayesB, BayesC, BayesL and BayesR methods for estimation of marker effects.
In order to compare the different methods, R software were used to simulate datasets. Simulation started with a base population of 100 animals, including 50 male and 50 female, which randomly mated for subsequent 50 generations. Generations were discrete and number of animals for each generation fixed at 100 animals. Thereafter, by randomly mating of animals in generation 50, reference population generation with 1000 animals were obtained and a threshold trait was simulated for each of them. Finally, Next validation population generation was produced by randomly mating of animals from the previous generation. Simulated genome for each animal consisted of 10 chromosomes with equal 1 Morgan lengths, each having 1000 evenly spaced SNPs. In each scenario QTLs were randomly distributed on genome and their substitution effects were drawn from one of five normal, uniform, t, gamma and Laplace distributions. Marker effects were estimated in reference population using five different Bayesian methods that differ with respect to assumptions regarding distribution of marker effects, including: Bayes A, Bayes B, Bayes C, Bayes L and Bayes R. These estimated markers effects were used for genomic breeding value predictions of animals in validation population which did not have any phenotype. The accuracy of the different methods was calculated as correlation between true and estimated genomic breeding values.
Results of this study showed that Bayes A, Bayes B and Bayes C predictions was affected by QTL numbers and distributions, while Bayes L and Bayes R had almost same accuracies for all scenarios. Scenario with gamma distribution for QTL effects and number of 10 QTLs had highest accuracy for Bayes A, Bayes B and Bayes C methods. Bayes C method was the best method when number of QTL was low. Increases in number of QTL from 10 to 150 were decreased accuracy of Bayes A, Bayes B and Bayes C methods and after that accuracy of different methods was constant and same. When distributions of QTL effects was uniform lower accuracy achieved rather than other distributions for same QTL number, instead distributions with unequal QTL variance like gamma increased accuracy of Bayes A, Bayes B, and Bayes C methods. These results were because of different extent of shrinkage of estimates of effects for different methods. Methods like Bayes B induced differential shrinkage of estimates relative to methods like Bayes R and Bayes L lead to higher accuracy for scenarios with low QTL number and unequal QTL variance.
Conclusion in a nutshell, different studied methods had diverse results for various QTL number and distribution, and its difficult to suggest a method that best in all situations. However, it is better to use methods like Bayes B when the number of QTLs was low and each QTL had not equal contribution to the trait of interest.