Comparing artificial-intelligence techniques with state-of-the-art parametric prediction models for predicting...

Area of Science:

• Agricultural Science
• Plant Breeding
• Genetics

Background:

• Soybean is a crucial crop for protein and oil, necessitating improved varieties.
• Genotyping offers a cost-effective alternative to phenotyping for crop breeding.
• Genomic prediction (GP) methods leverage marker data to predict crop performance.

Purpose of the Study:

• To compare the prediction accuracies of four GP methods in soybean.
• To evaluate methods including genomic best linear unbiased prediction (GBLUP), Gaussian kernel (GK), deep learning (DL), and arc-cosine kernel (AK).
• To assess method performance with and without genotype × environmental interaction (G×E) effects.

Main Methods:

• Utilized soybean nested association mapping data (1,379 genotypes, 6 environments).
• Compared prediction accuracies for grain yield, oil, and protein content.
• Evaluated GBLUP, GK, DL, and AK models, considering G×E effects.

Main Results:

• GBLUP consistently demonstrated superior prediction performance.
• GK and AK methods showed similar performance patterns to GBLUP.
• Deep learning (DL) generally performed worse, especially when G×E effects were included.

Conclusions:

• Conventional GBLUP remains a robust method for soybean genomic prediction.
• The performance of AI and kernel methods may require further optimization, particularly regarding hyperparameters and G×E interactions.
• Understanding method performance is key for efficient breeding of high-yield, high-quality soybean varieties.