Bayes U: A Genomic Prediction Method Based on the Horseshoe Prior

Abstract Text: We propose a novel method for genomic prediction, Bayes U, based on the Horseshoe prior. We compared it with other methods using simulations. All methods compared have different priors for their shrinkage profile. Evaluation of estimated SNP effects showed that Bayes U has stronger variable selection properties, assigning larger estimated effects to those SNPs with strong signals, and assigning more SNPs to have effects closer to zero. However, differences were less noticeable when assessing the accuracy of their overall prediction. Ridge regression and Bayesian Lasso have the lowest accuracies, but no differences were observed with Bayes U, Bayes A, Bayes B and Bayes C. Further studies are required to understand how these methods with different properties lead to similar predictions. The properties of Bayes U may prove to be a desirable behavior for QTL detection and may scale better for sequence data.

Keywords:

Genomic evaluation

Bayes U

Horseshoe prior