Comparing methods for approximating reliabilities in large-scale single-step genomic evaluations

Abstract

Accurate approximation of genomic estimated breeding value (GEBV) reliabilities is vital in single-step genomic prediction as reliable predictions of GEBV facilitate effective selection decisions. However, calculating exact reliabilities by inverting the left-hand side matrix of the mixed model equations is computationally infeasible for large datasets. In this study, we compared two approaches from Luke and Interbull for approximating genomic reliabilities for both genotyped and non-genotyped animals. The Luke approach uses effective record contributions (ERC) derived from the conventional EBV reliabilities as weights to approximate GEBV reliabilities for genotyped animals. A blended approach is used to implicitly account for residual polygenic (RPG) effects. Subsequently, genomic information is propagated to non-genotyped animals using ERC weights derived from the reliabilities of the genotyped animals. In contrast, the Interbull approach requires the derivation of a constant parameter, denoted , which is the genomic effective daughter contribution (EDC) gain via the Interbull GEBV test. This parameter is used to propagate genomic information to non-genotyped relatives through the pedigree. The final genomic reliabilities are obtained by combining conventional reliabilities with the genomic reliability gain. Notably, accuracy of reliabilities by this method highly depends on the precise estimation and regular updating of . In addition, this approach requires validation-based adjustments to correct inflated theoretical reliabilities observed in extremely large reference populations. In this study, both approaches were assessed and compared against exact reliabilities using a real dataset from the Finnish Red dairy population under a single trait model. The results demonstrated that the approximated reliabilities from both approaches were in close agreement with the exact reliabilities. Thus, both approaches can offer effective strategies for obtaining the reliabilities of GEBV in practical large-scale single-step evaluations.