Abstract:
|
Traditional selection for economically important quantitative traits in crop breeding is based on phenotypic records of individuals, which is long, expensive, and resource-dependent for crop improvement. With the availability of dense genome-wide marker data, genotypic value based selection is attractive because it increases the genetic gain rate and shortens the breeding cycle. Thus genomic prediction becomes increasingly important in crop breeding. In this study, we use the first generation of Oklahoma's dual-purpose 'Duster x Billings' breeding population to train statistical models for predicting grain yield. With the high-dimensional wheat data, we investigate parametric regression methods (e.g., Bayesian RR, Bayesian LASSO) and non-parametric models (e.g., RKHS,Random Forest) for prediction. The main objectives are to determine the best genomic prediction model specifically according to the prediction accuracy and biological relevance, and then to provide a proper choice of predictor markers for future breeding improvement. The selected best genomic prediction model will be used to assess the efficiency of genetic gain through two-generation validation.
|