A fast EM algorithm for genomic selection

Genomic selection is being adopted in many livestock breeding programs. Some industry applications use BLUP methods (called GS-BLUP) which are computationally fast but assume each marker effect is normally distributed with the same variance. The accuracy of prediction can often be increased by using models which not only allow marker variance to vary but also allow a large proportion of markers to have no effect. Meuwissen et al. (2001) called these methods BayesA and BayesB respectively. However implementing these Bayesian methods is computationally slow, particularly for large SNP panels. This paper gives details of an Expectation Maximisation (EM) algorithm called emBayesB for implementing a BayesB-like model which is both fast and accurate.