Enhancing genome-enabled prediction by bagging genomic BLUP.

Area of Science:

• Quantitative genetics
• Machine learning
• Statistical genomics

Background:

• Genomic best linear unbiased prediction (GBLUP) is a key tool for genomic prediction.
• Overfitting can limit GBLUP's predictive performance, especially with many markers and limited sample sizes.
• Resampling methods like bagging may improve model stability and accuracy.

Purpose of the Study:

• To investigate if bootstrap aggregating sampling (bagging) can enhance GBLUP's predictive ability.
• To adapt bagging for GBLUP at both genetic signal and marker effect levels.
• To evaluate bagging's performance in simulated and real-world genomic prediction scenarios.

Main Methods:

• Bagging was adapted for GBLUP, incorporating genetic signal and marker effects.
• Performance was assessed using four simulated datasets with varying quantitative trait loci (QTL) scenarios.
• Real-world data on wheat grain yield across four environments was analyzed.

Main Results:

• Bagging improved the predictive performance of GBLUP.
• Bagging enhanced GBLUP's robustness against overfitting.
• 25-50 bootstrap samples were sufficient for stable predictions and accurate mean squared error estimation.
• Theoretical reliabilities did not correlate with cross-validation accuracy.

Conclusions:

• Bagging is a viable method to improve GBLUP predictive accuracy and stability.
• The proposed metric for cross-validation uncertainty requires further assessment.
• Bagging offers a practical approach to mitigate overfitting in genomic prediction models.