A computationally efficient algorithm for genomic prediction using a Bayesian model.

Area of Science:

• Genomics
• Statistical Genetics
• Animal Breeding

Background:

• Genomic prediction using single nucleotide polymorphisms (SNPs) is crucial for livestock and crop breeding, and human disease risk assessment.
• Bayesian methods with non-linear SNP effect estimates offer high accuracy but are computationally intensive due to Markov chain Monte Carlo (MCMC) sampling.
• Existing computational demands limit the scalability of accurate genomic prediction models.

Purpose of the Study:

• To develop an efficient expectation-maximisation (EM) algorithm, emBayesR, for genomic prediction.
• To achieve accuracies comparable to the MCMC-based BayesR while substantially reducing computational time.
• To enable the application of advanced Bayesian genomic prediction to larger datasets.

Main Methods:

• Developed emBayesR, an approximate EM algorithm implementing the BayesR model with SNP effects from a mixture of normal distributions.
• Incorporated estimation of each SNP's effect while accounting for the error in estimating all other SNP effects.
• Compared emBayesR against BayesR using simulated data and real dairy cattle data (632,003 SNPs).

Main Results:

• emBayesR, incorporating error correction for SNP effect estimation, improved genomic prediction accuracy on both simulated and real data.
• Genomic prediction accuracy with emBayesR was only 0.5% lower than BayesR across nine dairy traits.
• emBayesR demonstrated up to an 8-fold reduction in computation time compared to BayesR.

Conclusions:

• The emBayesR algorithm achieves genomic prediction accuracies comparable to BayesR using simulated and real dairy SNP data.
• emBayesR offers a significant reduction in computational time, facilitating its use with larger genomic datasets.
• This efficient algorithm enhances the practical application of Bayesian methods in genomic selection.