Application and Optimisation of Genomic Selection for Wheat Breeding

Abstract

Plant breeding has a rich history of producing yield gains in bread wheat through the innovation and adoption of new technologies. This result is driven by extensive research, first in developing the technology, and second on establishing its application. Genomic selection is a recent technology which over the past decade has been the focus of extensive research effort. This research has been highly effective at developing the technology, and our attention should now pivot towards establishing and refining the parameters under which it should be applied. If genomic selection is to be successfully implemented in wheat breeding programmes breeders must be better informed on the optimal design of training strategies, and will also require cost-effective genotyping solutions. This body of work concentrates on delivering three overarching intended research outcomes: i) establish the achievable accuracy of genomic prediction in a large breeding population, ii) identify criteria for the optimal design of a genomic selection training strategy, and iii) investigate concepts and formulate methods for reducing the cost of implementing genomic selection. We present a dataset of unprecedented size in genomic selection studies, and utilise it to address these objectives. In the first component of the project we confirmed the significant potential of genomic selection by producing high prediction accuracies in a large and representative set of breeding germplasm, and showed genomic selection to be more accurate than marker assisted selection in all 14 traits tested. It was also demonstrated that genomic relationship information can be incorporated into the analysis of phenotype data to significantly improve model accuracy. The second component investigated factors affecting genomic prediction accuracy and how these relationships could be exploited in order to efficiently design accurate training strategies. We found that prediction accuracy continued to respond to training set size well beyond sizes previously tested in the literature, and that this response was independent of the genetic complexity of the trait. The impact of relatedness on prediction accuracy was highlighted, and it was shown that accuracy could be improved by increasing relatedness between training and prediction sets, or by increasing the diversity in the training set. To reduce the cost of implementing genomic selection, we present two novel methodologies for accurately utilising a low density genotyping platform. These approaches were shown to significantly increase the rate of genetic gain compared to a high density platform, with the same total genotyping expenditure. They could also be used to lower the cost of genomic selection without sacrificing genetic gain. The work presented here represents a significant resource which will inform pragmatic plant breeders on how to effectively and efficiently implement genomic selection in their programmes. The findings clarify uncertainties and overcome constraints associated with applying genomic selection, and can therefore be leveraged to facilitate increased rates of genetic gain in wheat breeding programmes around the world.