Genome-Wide Selection in Soft Winter Wheat: Effects of Training Population Size, Number of Markers, and Relatedness on Genomic Prediction Accuracy.

Genomic selection (GS) holds the promise of achieving higher genetic gains by using molecular markers as predictors of breeding values of individuals. The effect of training population (TP) size, marker number, and relatedness on the prediction accuracy of genomic selection for grain yield (GY), yield components, and agronomic traits in a diverse panel of soft winter wheat lines (N= 239) was evaluated by applying a standard single population cross-validation scheme under a ridge regression best linear unbiased prediction (RRBLUP) model. Prediction accuracies, r_GS ranged from -0.08 to 0.70 for measured traits. Increasing TP size resulted in an increase in r_GS, where optimum predictions reached when 80% of the lines were used as TP. Using subsets of markers (p <0.05) derived from association analyses also significantly increased r_GS among measured traits compared to using whole marker dataset. Relative efficiency of GS per year (RE_y) for GY increased from 0.98-3.71 to 1.60-5.90 when subsets of marker data were used. Using lines belonging to same subpopulation, Q to predict performance on the same group also had effects on r_GS values, particularly for lowly heritable trait such as GY, indicating the importance of relatedness between the training and validation populations to achieve optimal predictions. Additionally, using locations with high phenotypic correlations to predict line GY performance also showed effects on r_GS. Taken together, our results demonstrate the importance of TP size, relatedness, and marker number in the context of improving GS accuracies in soft winter wheat.