Genomic prediction for grain yield in a barley breeding program using genotype × environment interaction clusters

Genotype × environment interaction (GEI) is one of the key factors affecting breeding value estimation accuracy for agronomic traits in plant breeding. Measures of GEI include fitting prediction models with various kernels to capture the variance resulting from GEI, and characterizing trials into megaenvironment (ME) clusters within which breeding values can be estimated to remove the main GEI effects. However, many of the current approaches require observations of common genotypes across all trials, which is unavailable in most breeding programs. Our study introduces two methods that can be implemented on unbalanced data to categorize trials into clusters, where both need a correlation matrix between trials: one estimated via a factor analytic (FA) model and another estimated via weather variables. The methods were tested using empirical barley ( Hordeum vulgare L.) yield data in a commercial breeding program from 102 trials over 5 yr spread across multiple locations in Australia. Leave‐one‐year‐out cross‐validation achieved comparable predictive accuracies using either trials or clusters as the observed variable in GEI FA models (max. 0.45), which was higher than the accuracy achieved using the non‐GEI model (0.37). In the random cross‐validations, accuracies achieved within clusters (0.42–0.64) were mostly comparable with those achieved in the full population (0.62). In the within‐cluster validations, higher predictive accuracies were achieved when the training population was from the same cluster (mean 0.22) than outside of the cluster (mean 0.16). Our proposed methods of characterizing multienvironment trials into clusters provides a novel way to define training populations by reducing the variance resulting from GEI and could be implemented in any plant breeding program.