RESUMO
KEY MESSAGE: The optimization of training populations and the use of diagnostic markers as fixed effects increase the predictive ability of genomic prediction models in a cooperative wheat breeding panel. Plant breeding programs often have access to a large amount of historical data that is highly unbalanced, particularly across years. This study examined approaches to utilize these data sets as training populations to integrate genomic selection into existing pipelines. We used cross-validation to evaluate predictive ability in an unbalanced data set of 467 winter wheat (Triticum aestivum L.) genotypes evaluated in the Gulf Atlantic Wheat Nursery from 2008 to 2016. We evaluated the impact of different training population sizes and training population selection methods (Random, Clustering, PEVmean and PEVmean1) on predictive ability. We also evaluated inclusion of markers associated with major genes as fixed effects in prediction models for heading date, plant height, and resistance to powdery mildew (caused by Blumeria graminis f. sp. tritici). Increases in predictive ability as the size of the training population increased were more evident for Random and Clustering training population selection methods than for PEVmean and PEVmean1. The selection methods based on minimization of the prediction error variance (PEV) outperformed the Random and Clustering methods across all the population sizes. Major genes added as fixed effects always improved model predictive ability, with the greatest gains coming from combinations of multiple genes. Maximum predictabilities among all prediction methods were 0.64 for grain yield, 0.56 for test weight, 0.71 for heading date, 0.73 for plant height, and 0.60 for powdery mildew resistance. Our results demonstrate the utility of combining unbalanced phenotypic records with genome-wide SNP marker data for predicting the performance of untested genotypes.