Using cross-validation in a fast EM algorithm for genomic selection and complex trait prediction

The paper reports on changes to the EM algorithm emBayesB which estimates QTL effects using dense genome-wide SNP marker data. To overcome convergence issues, modifications were made to the original algorithm which included cross-validation for the estimation of model parameters. The modified algorithm called emBayesB_CV was used to analyse a trait simulated on real human genotypes consisting of 294,831 SNP measured on 3925 individuals. Three datasets were simulated for a trait determined by 10, 100 or 1000 additive QTL. The results showed that the modified algorithm emBayesB_CV was not only computationally fast, but also more accurate than GBLUP in predicting breeding value. However prediction accuracy declined as the size of QTL effects decreased due to the result that although emBayesB_CV could accurately locate the chromosomal location of large QTL effects, this was not the case for small QTL effects.