Tailoring AI and ML models for genotype-by-environment prediction leveraging environmental covariates: A European rye example.

key messagesWe explored target variable and loss function engineering for isolating genotype-by-environment interactions from main environmental effects to improve machine learning and deep neural network-based prediction of genotype performances using environmental covariates to make inference for untested locations and years. Accurate prediction of environment-specific genotype performance remains a major challenge in crop improvement and agricultural decision-making. Machine-learning (ML)- and deep neural network (DNN)-based predictions are often dominated by environmental main effects rather than genotype-by-environment interaction (GxE) effects, which hinders their adoption in plant breeding. This study aims to improve GxE predictive modeling in rye but also generally across crops by developing novel approaches to tailor ML and DNN models toward predicting environment-specific genotype differences and rankings rather than absolute performances. We introduce two methodologies: (1) target-variable engineering based on linear mixed-model decompositions to isolate GxE effects and (2) a custom-loss-function implementation of the mean squared error of differences (Piepho 1998) to optimize models directly for prediction of within-environment genotype differences. The motivating dataset covers major rye growing regions worldwide and models were evaluated using a comprehensive cross-validation (CV) framework. Our approaches improved predictive abilities of ML/DNN models by + 21 to + 62% compared to classical ML/DNN-based yield prediction. We increased predictive accuracy for environment-specific genotype rankings between + 15.0% and + 9.8% across different CV schemes over baseline genotypic main effects. Furthermore, historical weather records enabled prediction of genotype performances in future, untested years. This work demonstrates that tailored ML and DNN strategies can outperform classical methods for GxE prediction, offering practical value for plant breeders, growers and variety testing authorities.