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

Area of Science:

• Agricultural Science
• Computational Biology
• Genetics

Background:

• Accurate prediction of genotype performance across environments is crucial for crop improvement.
• Current machine learning (ML) and deep neural network (DNN) models often prioritize environmental main effects over genotype-by-environment interaction (GxE) effects, limiting their utility in plant breeding.

Purpose of the Study:

• To develop novel ML and DNN strategies for enhancing the prediction of environment-specific genotype differences and rankings.
• To improve the modeling of genotype-by-environment interactions (GxE) for better prediction of genotype performance in untested locations and years.

Main Methods:

• Implemented target-variable engineering using linear mixed-model decompositions to isolate GxE effects.
• Developed a custom loss function based on the mean squared error of differences to optimize models for predicting within-environment genotype differences.
• Evaluated models using a comprehensive cross-validation framework on a global rye dataset.

Main Results:

• Achieved significant improvements in ML/DNN predictive abilities, ranging from +21% to +62% compared to classical yield prediction methods.
• Increased predictive accuracy for environment-specific genotype rankings by +15.0% to +9.8% over baseline genotypic main effects.
• Enabled prediction of genotype performance in future, untested years using historical weather data.

Conclusions:

• Tailored ML and DNN strategies effectively outperform classical methods for GxE prediction.
• These advanced methods offer practical value for plant breeders, growers, and variety testing authorities in optimizing crop selection and management.
• The study demonstrates a significant advancement in predicting genotype performance under varying environmental conditions, crucial for agricultural decision-making.