EM algorithm for Bayesian estimation of genomic breeding values.

Area of Science:

• Genomics
• Quantitative Genetics
• Statistical Genetics

Background:

• Genomic selection models predict genome-wide breeding values (GBV) using numerous single nucleotide polymorphism (SNP) effects.
• Current Bayesian methods like Bayesian shrinkage regression (BSR) and stochastic search variable selection (SSVS) using Markov Chain Monte Carlo (MCMC) algorithms are computationally intensive.
• There is a need for efficient and accurate methods for practical genomic selection.

Purpose of the Study:

• To develop a computationally efficient Bayesian method for estimating SNP effects in genomic selection.
• To improve the prediction accuracy of genome-wide breeding values (GBV).
• To address the computational burden associated with MCMC-based Bayesian methods.

Main Methods:

• Developed an Expectation-Maximization (EM) algorithm applicable to BSR.
• Proposed a novel EM-based Bayesian method, weighted BSR (wBSR), which modifies BSR by incorporating SNP-specific weights based on trait association strength.
• Conducted simulation experiments to compare wBSR with MCMC-based BSR and SSVS.

Main Results:

• The EM-based wBSR method significantly reduced computational time compared to MCMC-based BSR.
• wBSR demonstrated improved accuracy in predicting GBV over MCMC-based BSR.
• However, wBSR's prediction accuracy was found to be inferior to MCMC-based SSVS, a current benchmark method.

Conclusions:

• The EM-based wBSR method offers substantial advantages in computational efficiency over MCMC-based Bayesian approaches.
• wBSR provides more accurate GBV predictions than MCMC-based BSR.
• wBSR is a practical and efficient method for genomic selection, especially when dealing with a large number of SNP markers.