Modelling and Using Spatial Effects in Nationwide Historical Data Improve Genomic Prediction of Rice Heading Date in Japan

Genomic prediction is a promising strategy for enhancing crop breeding efficiency. Historical data of breeding and cultivation tests from geographically wide regions presumably contain rich information for training genomic prediction models. Therefore, it is essential to explore methodologies to effectively handle such data. To improve the prediction accuracy of models using historical data, we incorporated a spatial model to account for spatial structures among field stations, in addition to conventional genomic prediction models. Targeting the rice heading date from historical data across Japan, we first constructed conventional genomic prediction models using genomic and/or meteorological elements as predictors. Next, we obtain the residual terms. Assuming that the residual terms were partly explained by the spatial effects assigned to each field station, a spatial model was applied to the residual terms and the spatial effects were calculated. Our genomic prediction models performed best when the genome, meteorological elements, and genome-meteorology interactions were included (model 3), and they performed second best when the genome and meteorological elements were included (model 2). For these genomic prediction models, residual terms were spatially biased and corrected for spatial effects. For the best model (model 3), the root mean squared errors (RMSE) of genomic prediction combined with spatial effects were approximately 3.6 days under tenfold cross-validation and approximately 5.1 days under leave-one-line-out cross-validation. The inclusion of the spatial effects improved the RMSEs by approximately 15% and 9% for the former and latter, respectively. Lines with highly improved predictions of the spatial effects were developed, mainly in the northern Tohoku region. The spatial effects were heterogeneous and regional patterns were detected. These findings imply that spatial effects are important not only for improving prediction performance but also for dissecting the model itself to identify the factors contributing to model improvement.