Multi-environment validation and efficiency of genomewide selection among maize doubled haploids

Abstract

Here I present two empirical studies on the effectiveness of genomewide selection within biparental populations in maize (Zea mays L.). To my knowledge, these are the first reported empirical studies on the effectiveness of genomewide selection compared to phenotypic selection, across multiple years and locations within individual biparental populations. In the first study, I report the results of a comparison of phenotypic selection and genomewide selection based on multiple years (2008-2012), multiple locations (six per year), and two testers. My objectives were to compare the accuracy of phenotypic selection, genomewide selection and an index for combined phenotypic and marker information. Phenotypic selection was always as accurate as or more accurate than genomewide selection. Selection based on marker and phenotypic information was slightly more accurate than genomewide selection or phenotypic selection alone. I concluded that for genomewide selection to be superior to phenotypic selection, the gains must not be measured in terms of the per-generation response with equal population sizes and selection intensities. In the second study I considered different total budgets and per-sample costs for phenotyping and genotyping in maize and assessed the observed and expected gains from the schemes for phenotypic selection and genomewide selection. My objectives were to determine whether or not genomewide selection is more efficient than phenotypic selection under a fixed budget and to give recommendations for implementing genomewide selection in a commercial maize breeding program. Whether or not genomewide selection was more efficient than phenotypic selection depended on the following factors: (1) accuracy of phenotypic selection and of genomewide selection for the trait in a given cross; (2) size of the training population; (3) total budget; (4) costs of producing and genotyping a DH line; and (5) number of selected lines. The results indicated that in general prediction accuracies above 0.50 usually lead to higher efficiency in genomewide selection. In general, the relative efficiencies decreased as size of the training population decreased. A larger total budget and lower costs of DH production and of genotyping would enable larger test populations, thus leading to higher efficiencies of genomewide selection.