Genomic Selection and Model Accuracy in Barley for Deoxynivalenol (DON)

Traditional marker assisted selection has been used in breeding over the last two decades to select individuals carrying favorable alleles at mapped QTLs. This approach is suited for traits controlled by QTLs with large effects, but is less effective for complex traits controlled by many QTLs with small effects.  In this study, we used six-row barley breeding lines that were evaluated over several years for several traits including deoxynivalenol (DON), a toxin produced by the fungal pathogen Fusarium graminearum.  A training population of 169 historical breeding lines that includes 134 parents used in the University of Minnesota barley breeding program was used to predict the performance of subsequent breeding lines developed between 2006 and 2010.  The goals of this study are to compare several genomic prediction models and determine if prediction accuracy changes over time.  We used a pedigree based prediction model and three marker-based prediction models (ridge regression, BayesCπ, and kernel models) to predict DON in breeding lines that were progeny of lines used in the training population.  We investigated several factors that could affect prediction accuracy including allele frequency, linkage disequilibrium, and genetic distance between the training population and selection candidates.