Genomic-Enabled Prediction in Maize Using Kernel Models with Genotype × Environment Interaction.

Area of Science:

• Plant breeding and genetics
• Quantitative genetics
• Agricultural science

Background:

• Multi-environment trials are crucial for selecting superior genotypes in plant breeding.
• Genomic prediction models aim to improve selection efficiency by utilizing marker data.
• Understanding genotype-by-environment (G×E) interactions is key for accurate predictions.

Purpose of the Study:

• To compare the prediction accuracy of four genomic-enabled prediction models.
• To evaluate the performance of linear (GBLUP) and nonlinear (Gaussian kernel) methods.
• To assess model performance across different traits and datasets in maize.

Main Methods:

• Fitted single-environment (SM), multi-environment (MM), multi-environment single variance G×E deviation (MDs), and multi-environment environment-specific variance G×E deviation (MDe) models.
• Utilized linear kernel Genomic Best Linear Unbiased Predictor (GBLUP) and nonlinear kernel Gaussian kernel (GK).
• Applied eight model-method combinations to Brazilian maize datasets (HEL and USP) for grain yield, plant height, and ear height.

Main Results:

• Multi-environment models with Gaussian kernel (MDe-GK, MDs-GK) showed the highest prediction accuracy.
• Gaussian kernel improved prediction accuracy over GBLUP, particularly for grain yield (up to 70% increase in USP dataset).
• Gains in prediction accuracy were smaller for plant and ear height and decreased with more challenging prediction scenarios.

Conclusions:

• Multi-environment models incorporating G×E interactions, especially with a Gaussian kernel, enhance prediction accuracy in maize breeding.
• The choice of model and kernel method significantly impacts prediction performance.
• These findings can guide the development of more effective genomic selection strategies.