School of Agriculture, Food and Ecosystem Sciences - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemAspects of statistical modelling for genomic selection( 2010)The research reported in this thesis investigated aspects of statistical models used for genomic selection. The importance of, and, interest in genomic selection is driven by the desire to increase the rate of genetic gain for commercially important traits. Genomic selection could increase the rate of genetic gain by increasing the accuracy of selection through the inclusion of DNA markers. Multiple methods and models have been proposed for implementing genomic selection. All methods have to overcome the problem that the number of DNA markers (p) is typically much larger than the number of phenotypic records (n) i.e. the p>n problem. One approach to this problem is to use Bayesian Inference which allows for an oversaturated model. Two simulation studies and a large data study were undertaken to gain a comprehensive understanding of what makes a robust and accurate Bayesian prediction model. Results from the simulation studies indicated that the match between the assumed QTL distribution and the true QTL distribution had an effect on the accuracy of the direct genomic values (DGV) produced by the different Bayesian models. Some of the models producing accurate DGV were computationally demanding. Subsequently, a novel Bayesian model using Stochastic Search Variable Selection (SSVS) for genomic selection was developed (Bayes SSVS). This model was demonstrated to produce accurate DGV and be computationally efficient. In contrast to the results from simulated studies, the results from a real dairy cattle data study showed a general equality in the accuracy of prediction across the various Bayesian models including Bayes SSVS. The exception was for traits with atypical genetic architectures such as fat percentage in milk where Bayes SSVS and other model selection approaches performed better than other approaches assuming that all markers equally contributed to the total genetic variation. The thesis also sought to explore the potential of genomic selection for improving novel traits that have been traditionally very difficult to select for. Energy Balance (EB) is a minimally recorded trait as the cost and measurement logistics mean it can only recorded on experimental farms. Using EB as a case study, it was demonstrated that genomic selection could provide the opportunity to select for EB and other minimally recorded through the accurate prediction of DGV. Additionally, selection for EB could be a valuable tool in finding a balance between production and non- production traits. Another attractive feature of some of the Bayesian models for genomic selection is they can be used to map QTL. Consequently, the establishment of significance when using multi-locus models for genome wide association studies was explored using a permutation testing approach. Three examples demonstrated that the permutation testing approach could correctly identify QTL. Two specialised approaches, permuting within strata, are presented. One approach accounted for a structured pedigree satisfying the condition of exchangeability. The second approach enabled the identification of secondary moderate QTL in the presence of a major QTL. The effect of the number of permutations needed was also examined; confirming previous results. This method was shown to provide accurate identification of QTL when compared with current approaches.