Genomic selection in biparental populations: assessment of parameters for optimum estimation set design

The prediction accuracy in genomic selection is affected by complex interactions of molecular, genetic and phenotypic factors. Despite the extensive use of biparental populations for empirical- and simulation-based studies, the highly variable prediction accuracies produced by cross-validation have not been further investigated. Understanding factors correlated with the variability in prediction accuracy could provide new insights for estimation set (ES) design. In this study, we employed simulations and show that parameters derived from marker data are not associated with prediction accuracy within biparental populations and, therefore, cannot serve as tool for optimum ES design. In contrast, the phenotypic variance in the ES is a major factor correlated with prediction accuracy. In particular, for estimation sets of small size, a large phenotypic variance probably ensures that more quantitative trait loci are segregating in the ES which consequently allows a better marker effect estimation. While the ES phenotypic variance is not known beforehand, we discuss approaches how prediction accuracy could nevertheless be maximized for small estimation sets, towards a more efficient implementation of genomic selection within biparental families in plant breeding.