Improving accuracy of genomic prediction in dairy and beef cattle
The overall goal of this thesis was to improve the accuracy of genomic prediction in dairy and beef cattle by developing, evaluating and enhancing novel or existent models and approaches for genomic selection. Four studies were conducted to fulfill this goal. In the first study, the impact of using genotypes imputed from low density panels for genomic prediction was evaluated and compared between a Bayesian mixture model and the Genomic Best Linear Unbiased Prediction (GBLUP) method. Results showed that for traits affected by a few large QTL, the Bayesian mixture model resulted in greater reduction in accuracy of genomic prediction, compared to GBLUP. However, for all SNP panels, scenarios and all traits studied, the Bayesian mixture model produced greater or similar accuracy, compared to the GBLUP method. In the second study, a new computing algorithm, called right-hand side updating strategy (RHSU), was proposed and compared to the conventional Gauss-Seidel residual update algorithm (GSRU) for genomic prediction. Results showed that RHSU would outperform GSRU once the sample size exceeded a fraction of the number of the SNPs. As the sample size continued to grow, the RHSU algorithm became more efficient than GSRU. In the third study, three different strategies of forming a training population for genomic prediction, within-breed, across-breed and pooling data from different breeds, were evaluated in Angus and Charolais steers using phenotypes on residual feed intake (RFI) and genotypes on the Illumina BovineSNP50 Beadchip (50k). Results suggested that using the 50k SNP panel, within-breed genomic prediction was a safe strategy; across-breed prediction resulted in the lowest accuracy; pooling data from different breeds had a potential to improve the accuracy but should be conducted with caution due to possible loss of accuracy. In the last study, a multi-task Bayesian learning model was proposed for multi-population genomic prediction. The performance of the multi-task model was evaluated in Holstein and Ayrshire dairy breeds. Results showed that the multi-task Bayesian learning model is effective and could be beneficial to smaller populations where only a limited number of training animals are available.