A Bayesian method and its variational approximation for prediction of genomic breeding values in multiple traits

Area of Science:

• Quantitative Genetics
• Animal Breeding
• Plant Breeding

Background:

• Genomic selection (GS) predicts breeding values without phenotypes, but typically focuses on single traits.
• Breeding programs often target multiple correlated traits, necessitating multi-trait GS for improved accuracy.
• Current multi-trait GS models face high computational demands, requiring efficient prediction methods.

Purpose of the Study:

• To develop and evaluate a multi-trait genomic selection model for joint prediction of genomic breeding values (GBVs).
• To compare the accuracy of multi-trait versus single-trait GBV prediction.
• To assess computational efficiency of different Bayesian estimation methods for multi-trait models.

Main Methods:

• A Bayesian regression model with variable selection was developed for joint multi-trait GBV prediction.
• Markov Chain Monte Carlo (MCMC) iteration (MCBayes) and variational approximation (varBayes) were devised for parameter estimation.
• Simulated datasets with SNP genotypes and phenotypes for three traits were used for comparison.

Main Results:

• Multi-trait analysis significantly improved GBV prediction accuracy for low-heritability traits correlated with high-heritability traits.
• Prediction accuracy for uncorrelated low-heritability traits was comparable or lower with multi-trait analysis.
• varBayes offered substantial computational time reduction compared to MCBayes, with only a slight decrease in prediction accuracy.

Conclusions:

• Multi-trait genomic selection is more beneficial than single-trait analysis for correlated traits.
• varBayes provides a computationally efficient and practical alternative to MCBayes for multi-trait genomic selection.
• The computational advantages of varBayes outweigh minor accuracy losses, making it suitable for practical breeding applications.