Structured Multimodal Deep Learning improves Genomic Prediction in Future Environments

Area of Science:

• Genetics
• Plant Breeding
• Machine Learning
• Bioinformatics

Background:

• Predicting crop phenotypes from genetic and environmental data is crucial for plant breeding.
• Deep neural networks (DNNs) show promise for modeling complex biological processes but often underperform linear models.
• Existing DNNs struggle with issues like greedy learning and over-reliance on single data types, limiting their effectiveness in genotype-by-environment interaction prediction.

Purpose of the Study:

• To develop a novel deep learning framework, the Structured Interaction Neural Network (SINN), for accurate phenotype prediction.
• To address the limitations of current deep learning methods in capturing genotype-by-environment interactions.
• To improve genomic prediction accuracy for crop yield in new environments and genotypes.

Main Methods:

• Developed SINN, integrating statistical decomposition of genetic, environmental, and interaction effects with DNNs.
• Dissected phenotype prediction into isolated component modeling tasks to identify generalization limitations.
• Evaluated SINN on two large-scale maize multi-environment trial datasets (Genomes to Fields and MaizeGEP).

Main Results:

• SINN demonstrated superior yield prediction accuracy compared to traditional BLUP-based methods and previous DNNs on both datasets.
• On the Genomes to Fields dataset, SINN achieved higher accuracy (0.63 vs. 0.43) and lower RMSE (2.40 vs. 2.46 Mg/ha) than benchmarks.
• On the MaizeGEP dataset, SINN also outperformed benchmarks with higher accuracy (0.79 vs. 0.49).

Conclusions:

• SINN effectively combines statistical genetics principles with deep learning for enhanced genomic prediction.
• The model's modular approach improves the understanding of prediction limitations, particularly generalization to new environments.
• SINN offers an accurate, modular, and scalable solution for genomic prediction in diverse agricultural settings.