Multitrait machine- and deep-learning models for genomic selection using spectral information in a wheat breeding...

Area of Science:

• Plant breeding
• Genomics
• Artificial Intelligence

Background:

• Genomic selection (GS) revolutionized plant breeding value prediction.
• Increasing data and slow genetic gain necessitate AI exploration.
• Machine and deep learning show promise for complex trait prediction.

Purpose of the Study:

• Optimize multitrait (MT) machine and deep learning models for wheat grain yield and protein content prediction.
• Compare MT-GS models with unitrait (UT) models and traditional methods (GBLUP, Bayesian).
• Utilize spectral information for enhanced prediction accuracy.

Main Methods:

• Developed and compared four machine/deep learning UT and MT models.
• Evaluated models using 650 spring wheat recombinant inbred lines (RILs) over three years.
• Collected spectral data at heading and grain filling stages.

Main Results:

• MT-GS models outperformed UT-GS models by 0-28.5% and -0.04-15%.
• Random Forest and Multilayer Perceptron were the top-performing ML/DL models.
• Bayesian models had lower accuracy and higher computational cost than ML/DL models.
• Green normalized difference vegetation index (GNDVI) was a key predictor for grain protein content.

Conclusions:

• Machine and deep learning-based MT-GS models enhance prediction accuracy for wheat breeding.
• These AI-driven models are suitable for large-scale breeding programs.
• Spectral information integration improves trait prediction.