Optimal designs in genomic selection

Recently, many plant and animal breeders have been using genome-wide genetic markers and statistical methods to aid with selection of genetic material. These methods, termed genomic selection (GS), make selections based on estimates of breeding values obtained from a prediction model computed from phenotypic and genomic data of a training population. The precision of the predictions strongly depends on the genetic diversity of the training population (TP). The objectives of this research were (1) To present a new method for creating a TP that maximizes genetic diversity using either the most important genomic markers or the first few principal components (PCs) of the genomic data as inputs into A, D, and V optimal design algorithms; and (2) To evaluate the average predictive ability of the A, D, and V optimal TPs and compare their predictabilities with TPs based on random sampling, the commonly used approach. Using data from the University of Nebraska red winter wheat breeding program, results showed that when created the TP using either the most significant markers or the first PCs, the gain of the average predictive ability was higher in all optimal designs compared with random sampling with an average increase by 13.425% over random sampling. In addition, it was estimated that genetic gains of selection can be increased by 2.8348 and 3.3538 times when using the p1 significant markers and the first p1 PCs compared with the genetic gain of 1.8306 random sampling TP, respectively