Biological Prior Knowledge-Embedded Deep Neural Network for Plant Genomic Prediction

Area of Science:

• Genomics
• Plant Breeding
• Artificial Intelligence

Background:

• Genomic prediction accelerates plant trait improvement using genotypic data.
• Traditional methods (GBLUP, SVR) struggle with high-dimensional data and nonlinearities.
• Deep learning offers potential for enhanced genomic prediction.

Purpose of the Study:

• Introduce iADEP, a deep learning model for genomic prediction.
• Integrate additive, dominant, and epistatic genetic effects.
• Incorporate biological prior knowledge for improved accuracy.

Main Methods:

• Developed iADEP, a deep learning model.
• Fused SNP data with genetic interactions and GWAS results via an SNP embedding block.
• Employed a multi-head attention mechanism to integrate local and global decoders, including omics data.

Main Results:

• iADEP outperformed existing methods in genotype-to-phenotype prediction across four datasets.
• Ablation studies validated the effectiveness of the SNP embedding.
• Demonstrated iADEP's flexibility in combining and fusing other omics data.
• Full SNP sets generally yielded optimal performance, and transductive vs. inductive learning differences were explored.

Conclusions:

• iADEP offers a novel approach for AI-driven plant breeding.
• The model effectively integrates biological prior knowledge.
• iADEP facilitates the combination of diverse omics data for enhanced prediction.