A Stacking Ensemble Learning Framework for Genomic Prediction.

Area of Science:

• Genomics
• Machine Learning
• Animal Breeding

Background:

• Genomic datasets are large and complex, requiring advanced analytical tools.
• Single machine learning methods often show unsatisfactory performance in genomic selection.
• Accurate prediction of genomic estimated breeding values (GEBVs) is crucial for genetic improvement.

Purpose of the Study:

• To develop and evaluate a stacking ensemble learning framework (SELF) for enhanced genomic prediction.
• To compare the prediction accuracy of SELF against established methods (GBLUP, BayesB) and its base learners (SVR, KRR, ENET).
• To assess the robustness of SELF across diverse genetic architectures in real datasets.

Main Methods:

• Construction of a stacking ensemble learning framework (SELF) integrating three machine learning methods.
• Analysis of three real-world genomic datasets with varying genetic architectures.
• Comparative evaluation of prediction accuracy using SELF, base learners, GBLUP, and BayesB.

Main Results:

• SELF consistently outperformed its individual base learners (SVR, KRR, ENET) in prediction accuracy.
• SELF demonstrated an average 7.70% higher prediction accuracy than GBLUP across the three datasets.
• SELF showed greater robustness than BayesB for most traits, with a notable exception for milk fat percentage in German Holstein cattle.

Conclusions:

• The developed stacking ensemble learning framework (SELF) significantly enhances genomic prediction accuracy.
• SELF offers a more robust and accurate alternative to existing methods for predicting GEBVs.
• SELF shows strong potential for application in genomic selection for both animals and plants.