Training Population Selection and Use of Fixed Covariates to Optimize Genomic Predictions in a Historical Southeastern USA Winter Wheat Panel.

In the context of genomic selection the use of historical data from unbalanced breeding nurseries as training populations offers the possibility to integrate genomic predictions into the existing pipeline of breeding programs. However, the optimal approach to the use of historical unbalanced data for genomic selection in commercial breeding programs is uncertain. We used cross-validation to evaluate the predictive ability of genomic predictions in a set of 467 genotypes from the Gulf Atlantic Wheat Nursery (GAWN) evaluated from 2008 to 2016. We evaluated the impact on predictive ability of different training population sizes and selection methods (Random, Clustering, PEVmean and PEVmean1) and the effect of inclusion of major QTL for heading date, plant height and powdery mildew resistance as fixed covariates. Increases in predictive ability as the size of the training population increased were more evident for the Random and Clustering selection methods and reached a maximum at the largest population size of 350 individuals. Selection methods based on minimization of the prediction error variance outperformed the other methods evaluated across all the population sizes with an optimum of 200 to 250 individuals. Major gene covariates always improved prediction accuracy, with the greatest increases in performance from the use of multiple covariates in combination. Maximum prediction abilities after optimization of the training population were 0.64, 0.56, 0.71, 0.73, 0.60 for grain yield, test weight, heading date, plant height and powdery mildew resistance, respectively. Our results demonstrate the potential value of combining these unbalanced sets of phenotypic records with genome wide marker data for prediction of untested genotypes in collaborative breeding efforts in soft red winter wheat.