Ensemble of Bayesian alphabets via constraint weight optimization strategy improves genomic prediction accuracy

Area of Science:

• Quantitative genetics
• Bioinformatics
• Plant breeding

Background:

• Genomic prediction models are crucial for accelerating crop improvement.
• Existing models often have limitations in prediction accuracy.
• Ensemble methods offer a potential solution by combining multiple models.

Purpose of the Study:

• To develop and evaluate a weight optimization-based ensemble framework, EnBayes, for improved genomic prediction accuracy.
• To compare EnBayes performance against individual Bayesian models, meta-learning approaches, and traditional genomic prediction and machine learning models.
• To investigate the impact of objective functions and model selection on ensemble accuracy.

Main Methods:

• Incorporation of eight Bayesian models (BayesA, BayesB, BayesC, BayesBpi, BayesCpi, BayesR, BayesL, BayesRR) into an ensemble framework.
• Optimization of model weights using a genetic algorithm.
• Evaluation on 18 diverse crop datasets.
• Exploration of new objective functions for Pearson's correlation coefficient and mean square error.
• Comparison with meta-learning (random forest, quantile regression forest, ridge regression) and other models (GBLUP, rrBLUP, SVM, RF, XGBoost, LGBoost).

Main Results:

• EnBayes demonstrated superior prediction accuracy compared to individual Bayesian models across 18 crop datasets.
• The ensemble model's accuracy was influenced by the number and performance of constituent models.
• EnBayes outperformed meta-learning approaches and traditional genomic prediction/machine learning models.
• Proposed objective functions improved prediction accuracy metrics.

Conclusions:

• The EnBayes framework significantly enhances genomic prediction accuracy.
• Weight optimization and careful model selection are key to successful ensemble prediction.
• EnBayes represents a valuable tool for advancing genomic selection in crop breeding programs.