Predicting GxE in Biomass and Yield of Rice: A New Modelling Method Unifying Crop Models and Genomic Prediction Models

While the recent development of genomic prediction (GP) for agronomical traits seems remarkable, prediction of yield still has room of improvement due to the difficulty of modelling of environmental effect. In this study, we tried to construct prediction models for biomass and panicle weight of rice in which both genotypic and environmental effects were considered.

We cultivated two parental varieties, ‘Kinmaze’ and ‘Koshihikari’, and their 123 RILs in 2014-2015 in an experimental field in Tsukuba, Japan. The RILs were genotyped with 362 SNPs.

We developed a two-step prediction model. In the first step, growth-related traits with low sensitivity to environmental effect were predicted by simple GP. Note that heading date was predicted by using DVR (developmental rate) model based on daily temperature and day length. In the second step, target traits were predicted based on the growth-related traits either with a crop growth (CG) model or with a machine-learning (ML) model. In CG, target traits were predicted as the summation of daily growth, which was calculated from growth-related traits and environmental information. In ML, target traits were predicted from the growth-related traits either with multi-linear regression or with random forest regression.

As a result, the model with CG worked better than the simple GP in biomass for both known and unknown environments, although prediction accuracy was still low in panicle weight by using all models.

We conclude that an efficient use of environmental information will enhance the accuracy and robustness of genomic prediction in yield-related traits influenced by environmental fluctuations.