Accurate predictions of barley phenotypes using genomewide markers and environmental covariates

Predicting the performance of new plant genotypes under new environmental conditions could accelerate the development of locally adapted and climate resilient cultivars. Enabling these predictions may rely on extending the genomewide prediction framework to include environmental covariates (EC), but such models have generally been tested under limited, less breeding-realistic contexts. Using a barley (Hordeum vulgare L.) multi-environment dataset, our objectives were to compare multi-environment prediction models and scenarios that target genotypes from different breeding generations, use different levels and timescales of ECs, and are applied to different agronomic and quality traits. When predicting the phenotypes of previously tested genotypes in untested environments, models that included the interaction of genomewide markers and pre-selected in-season ECs resulted in more accurate predictions (r MG or r MP) within (r MG = 0.56-0.94) and across (r MP = 0.63-0.87) environments; similar accuracy was achieved within (r MP = 0.46-0.89) and across (r MP = 0.87-0.95) locations when using only ECs from realistically available historical climate data. Shifting the prediction target to a distinct, untested offspring population slightly reduced model performance within environments or locations, but r MP across environments (r MP = 0.60-0.86) or locations (r MP = 0.87-0.94) remained very high. Though we achieved moderately high r MP for most traits in the challenging scenario of predicting the offspring population in holdout environments, the similarity between training and target environments, like that between populations, will be a limiting factor for enabling accurate predictions of new genotypes under new growing conditions.

Automated summary

Extending the genomewide prediction framework to include environmental covariates can improve the accuracy of predicting the performance of new plant genotypes under new environmental conditions. However, the similarity between training and target environments is a limiting factor for accurate predictions.