Implications of using genomic prediction within a high-density SNP dataset to predict DUS traits in barley

Alternative methods for genomic prediction of traits and trait differences are compared and recommendations made. We make recommendations for implementing methods in the context of DUS testing. High-throughput genotyping provides an opportunity to explore the application of genotypes in predicting plant phenotypes. We use a genome-wide prediction model to estimate the contribution of all loci and sum over multiple minor effects to predict traits. A potential use is in plant variety protection to discriminate among varieties on distinctness. We investigate this use with alternate scenarios in a set of 431 winter and spring barley varieties, with trait data from UK DUS trials comprising 28 characteristics, together with SNP genotype data. Firstly, each trait is predicted from genotypes by ridge regression with discrimination among varieties using predicted traits. Secondly, squared trait differences between each pair of varieties are regressed on genetic distances between each variety by ridge regression, with discrimination among varieties using the predicted squared trait differences directly. This latter approach is analogous to the use of phenotype and marker differences introduced to human genetic linkage analysis by Haseman and Elston and to the analysis of heritability in natural populations of plants by Ritland. We compare correlations between methods, both trait by trait and summarised across all traits. Our results show wide variation among correlations for each trait. However, the aggregate distances calculated from values predicted by genotypes show higher correlations with distances calculated from measured values than any previously reported. We discuss the applicability of these results to implementation of UPOV Model 2 in DUS testing and suggest 'success criteria' that should be considered by testing authorities seeking to implement UPOV Model 2.