Exploring extensions and properties of expectation-maximization methods for whole genome prediction

Abstract Text:

As the density of single nucleotide polymorphism (SNP) marker panels increase, computational efficiency becomes a greater consideration for whole genome prediction (WGP) such that algorithms other than Markov Chain Monte Carlo (MCMC) might be more important. One popular alternative is the expectation maximization (EM) algorithm. Our group has previously extended BayesA to an anteBayesA model using inference based on MCMC.  We explore an EM analogue of this anteBayesA model called EM-anteBayesA and compare it to a more conventional EM-BayesA strategy developed earlier.  By both simulation and and application to the heterogeneous stock mice dataset, we find EM-anteBayesA has comparable accuracy to its MCMC analogue.  Furthermore, we demonstrate that it is feasible to estimate key hyperparameters in EM-(ante)BayesA models.  However, we also discovered that accuracies of genomic breeding values using these EM-based methods may depend on starting values for SNP effects. 

Keywords:

Computational efficiency

Genomic selection

EM