Single‐Step Reaction Norm Models for Genomic Prediction in Multienvironment Recurrent Selection Trials

Abstract

In recurrent selection programs, progeny testing is done in multienvironment trials, which generates genotype × environment interaction (G × E). Therefore, modeling G × E is essential for genomic prediction in the context of recurrent genomic selection (RGS). Developing single‐step, best linear unbiased prediction‐based reaction norm models (termed RN‐HBLUP) using data from nongenotyped and genotyped progenies, can enhance predictive accuracy. Our objectives were to evaluate: (i) a class of RN‐HBLUP models accommodating combined relationship of pedigree and genomic data, environmental covariates, and their interactions for prediction of phenotypic responses; (ii) the predictive accuracy of these models and the relative importance of main effects and interaction components; and (iii) the influence of different grouping strategies of genetic–environmental data (within selection cycles or across cycles) on prediction accuracy of the merit for untested progenies. The genetic material comprised 667 S1:3 progenies of irrigated rice (Oryza sativa L.) and six check cultivars. These materials were evaluated in yield trials conducted in 10 environments during three selection cycles. Genomic information was derived from single‐nucleotide polymorphism markers genotyped on 174 progenies in the third cycle. We evaluated six predictive models. Environmental covariates and G × E interaction explained a significant portion of the phenotypic variance, increasing accuracy and decreasing the bias of phenotypic prediction. Within‐cycle data were sufficient for accurate prediction of untested progenies, even in untested environments. We concluded that the RN‐HBLUP model, with the comprehensive structure, could be useful in improving the prediction accuracy of quantitative traits in RGS programs.