Genomic Prediction Using Bayesian Regression Models With Global-Local Prior.

Area of Science:

• Genomics and Bioinformatics
• Statistical Genetics
• Machine Learning in Biology

Background:

• Bayesian regression models are crucial for genomic prediction in diverse species.
• Existing models face challenges in effectively handling marker effect shrinkage and selection.
• Global-local priors offer a promising approach to improve genomic prediction accuracy.

Purpose of the Study:

• To introduce two novel Bayesian models, BayesHP and BayesHE, incorporating global-local priors for genomic prediction.
• To compare the performance of BayesHP and BayesHE against established genomic prediction models using simulated and real data.
• To evaluate the adaptability of these models to various genetic architectures and trait types.

Main Methods:

• Developed BayesHP using a Horseshoe+ prior and BayesHE using a half-t distribution for the local parameter.
• Compared BayesHP and BayesHE with GBLUP, BayesA, BayesB, and BayesU models.
• Utilized simulated data and real datasets from cattle (milk production, health, type traits) and mice (type, growth traits) for performance assessment.

Main Results:

• Global-local prior models (BayesU, BayesHP, BayesHE) outperformed others for traits with high heritability and few quantitative trait loci (QTL).
• BayesHE showed optimal or near-optimal performance across all tested traits, highlighting its adaptability.
• BayesHP was less effective than classical models for most traits, suggesting its suitability for traits with major QTL.

Conclusions:

• BayesHE's automatic hyperparameter estimation enhances its adaptability for diverse genomic prediction tasks.
• Auto-estimating the degree of freedom (as in BayesHE) is preferable to increasing local parameter layers (as in BayesHP).
• The introduction of global-local priors with unknown hyperparameters opens new avenues for developing advanced Bayesian genomic prediction models.