Yield prediction through integration of genetic, environment, and management data through deep learning

Area of Science:

• Agricultural Science
• Computational Biology
• Genetics

Background:

• Predicting phenotypic outcomes from genotype, environment, and management is crucial for agriculture, research, and conservation.
• Advancements in computational methods are expanding the possibilities for accurate phenotypic prediction.

Purpose of the Study:

• To predict maize yield using deep neural networks (DNNs).
• To compare DNNs with conventional linear and machine learning models.
• To evaluate the impact of incorporating interactions between different data types.

Main Methods:

• Deep neural networks were employed for maize yield prediction.
• Model performance was compared using best linear unbiased predictor (BLUP) and other machine learning approaches.
• The study investigated the effects of optimizing DNN submodules versus optimizing the entire model at once.
• Interactions between disparate data types (genotype, environment, management) were examined.

Main Results:

• Deep learning and BLUP models incorporating data interactions demonstrated superior performance.
• BLUP models yielded the lowest average error, while DNN models exhibited greater consistency.
• Optimizing DNN submodules for individual data types improved performance compared to a holistic optimization approach.
• Including interactions reduced the importance of weather and management features at extreme seasonal time points.

Conclusions:

• Deep learning presents a powerful tool for predicting complex traits in intricate environments.
• The study highlights the significance of data interactions in improving prediction accuracy.
• DNNs offer a potential mechanism for a deeper understanding of environmental and genetic factor influences on phenotypes.