Using Environmental Similarities To Design Training Sets For Genomewide Selection

In plant breeding, the goal of genomewide selection is to predict the merit of unobserved individuals, particularly those in the next breeding generation. Predictions of these individuals in unobserved or future environments would be of additional use to a breeder. For many of the complex traits targeted in breeding, this may require management of genotype x environment interactions by, for example, using data from homogeneous groups of environments. Our objectives were to assess the accuracy of genomewide predictions in unobserved environments both within and between breeding generations; we aimed to compare training sets that included data from all possible environments with those that included data from (a) decreasingly similar environments or (b) discrete clusters of similar environments. A 183-line spring barley (Hordeum vulgare L.) training population and 50-line offspring test population were phenotyped in 29 location-year environments for grain yield, heading date, and plant height. Environmental similarities were measured using phenotypic data, geographic distance, or environmental covariables. When using training data from more, but decreasingly similar environments, prediction accuracy increased, but marginal gains declined; in some cases, accuracy declined with additional data. Clusters of environments informed by phenotypes (i.e., phenotypic correlations or multiplicative models) typically improved prediction accuracy within a generation, but not between generations (offspring population). Our study suggests that, as an alternative to using data from all available environments, informative subsets may be advantageous for genomewide predictions within a single breeding generation, but not between generations.