A Multi-Trait Gaussian Kernel Genomic Prediction Model under Three Tunning Strategies

Area of Science:

• Plant breeding
• Genomics
• Statistical genetics

Background:

• Genomic selection (GS) has the potential to revolutionize plant breeding but faces challenges in practical implementation, primarily due to factors affecting prediction accuracy.
• The choice of statistical machine learning methods and their tuning significantly impacts the success of GS models.

Purpose of the Study:

• To explore the impact of different tuning methods on prediction accuracy within a multi-trait framework.
• To compare manual tuning, grid search, and Bayesian optimization for the Gaussian kernel with a multi-trait Bayesian Best Linear Unbiased Predictor (GBLUP) model.

Main Methods:

• Utilized a multi-trait GBLUP model with a Gaussian kernel.
• Implemented and compared three tuning strategies: manual tuning, grid search, and Bayesian optimization.
• Evaluated performance across 5 real datasets from plant breeding programs.

Main Results:

• Grid search and Bayesian optimization improved prediction accuracy by 1.9% to 6.8% compared to manual tuning.
• The careful tuning of GS models is essential for enhancing prediction accuracy, even with marginal improvements.
• Advanced tuning methods require greater computational resources but yield better results.

Conclusions:

• Automated tuning methods (grid search, Bayesian optimization) are superior to manual tuning for improving GS accuracy in multi-trait scenarios.
• Careful parameter tuning is a critical step for maximizing the effectiveness of genomic selection in plant breeding.
• Investing in computational resources for advanced tuning methods is justified by the gains in prediction accuracy.