Multimodal deep learning methods enhance genomic prediction of wheat breeding

Area of Science:

• Plant breeding
• Genomics
• Machine learning

Background:

• Genomic prediction (GP) models assess unobserved phenotypes in plant breeding.
• Few methods link genomics with phenomics (imaging) or genotype-environment interactions (GE).
• Deep learning (DL) shows potential for GP accuracy but hasn't been applied to linked genomics-phenomics data.

Purpose of the Study:

• To compare a novel deep learning (DL) method with conventional GP models.
• To evaluate DL for genomic prediction accuracy using linked genomics and phenomics data.
• To assess DL's performance in accounting for genotype-environment interactions (GE).

Main Methods:

• Utilized two wheat datasets (DS1 and DS2).
• Compared DL against GBLUP, GBM, and SVR models.
• Applied DL to predict phenotypes using genomic and phenomic data under different environmental conditions.

Main Results:

• DL achieved higher genomic prediction accuracy than GBLUP when predicting irrigated environments from drought data.
• DL showed competitive or slightly superior accuracy compared to GBLUP in other scenarios.
• The novel DL method demonstrated strong generalization capabilities for multi-input data.

Conclusions:

• The novel DL method offers a promising approach for enhancing genomic prediction accuracy in plant breeding.
• DL effectively integrates genomics and phenomics, improving predictions under varying environmental conditions.
• The DL method's modular design allows for future enhancements and broader applications in multi-input data analysis.