Genomic prediction for targeted populations of environments in oat (Avena sativa)

Context Long-term multi-environment trials (METs) could improve genomic prediction models for plant breeding programs by better representing the target population of environments (TPE). However, METs are generally highly unbalanced because genotypes are routinely dropped from trials after a few years. Furthermore, in the presence of genotype × environment interaction (GEI), selection of the environments to include in a prediction set becomes critical to represent specific TPEs. Aims The goals of this study were to compare strategies for modelling GEI in genomic prediction, using large METs from oat (Avena sativa L.) breeding programs in the Midwest United States, and to develop a variety decision tool for farmers and plant breeders. Methods The performance of genotypes in TPEs was predicted by using different strategies for handling GEI in genomic prediction models including systematic and/or random GEI components. These strategies were also used to build the variety decision tool for farmers. Key results Genomic prediction for unknown genotypes, locations and years within TPEs had moderate to high predictive ability, accuracy and reliability. Modelling GEI was beneficial in small, but not in large, mega-environments. The latest 3 years were highly predictive of performance in an upcoming year for most years but not for years with unusual weather patterns. High predictive ability, accuracy and reliability were obtained when large datasets were used in TPEs. Conclusions Deployment of historical datasets can be accomplished through meaningful delineation and prediction for TPEs. Implications We have shown the performance of a simple modelling strategy for handling prediction for TPEs when deploying large historical datasets.