Genomic Selection Accuracy using Historical Data Generated in a Wheat Breeding Program

Abstract

Cross-validation (CRV) methods were designed to simulate genomic selection (GS) for yield in a wheat ( L.) breeding program with data of 318 genotypes grown over an 11-yr period at six locations in France. Two methods, CVSWO (cross-validation-specific without location as factor) and CVSW (cross-validation-specific with location as factor), included 11 folds, each comprising genotypes grown during a specific year and each representing target populations, while the remaining folds comprising genotypes grown during the other 10 yr represented training populations. These methods were compared with CVRWO (cross-validation-random without location as factor) and CVRW (cross-validation-random with location as factor), designed to simulate standard CRV while retaining the structure of the first two CRV methods; the same 318 genotypes were used to create 11 folds, each comprising randomly selected genotypes. Results suggest the accuracy of the CRV methods using specifically selected genotypes (correlation coefficient between (marker based) estimate of breeding value and observed phenotype [] = 0.20) based on years grown were significantly less than methods using randomly selected genotypes ( = 0.40–0.50). These results imply wheat yield is more difficult to predict for unknown, futuristic years than standard CRV methods suggest. An alternative measure of accuracy based on predicted genotypic ranks, termed predicted rank conversion (PRC), was implemented for the purpose of improving accuracies and reducing the differences between CRV methods.