Optimizing the Training Population Composition to Improve Genomic Prediction Accuracy across Selection Cycles of a Barley Breeding Population

Genomic selection (GS) is becoming routinely used in plant breeding. To date, the literature and anecdotal experience of users indicates that, while it works, the accuracy of GS is greatly influenced by the composition of the training population (TP). The University of Minnesota barley breeding program has been using and evaluating GS since 2009. Now, a balanced experimental dataset including 50 progeny selected via GS and 50 progeny selected at random from each of the first three cycles of selection has been evaluated in multiple trials for grain yield and deoxynivalenol concentration, the toxin produced by the fungal pathogen Fusarium graminearium. Using this dataset to estimate the true breeding values of selection candidates in each breeding cycle, we retroactively assess the effect that three TP updating and optimization strategies have on prediction accuracy, which include i) amending the original TP to include breeding lines to more comprehensively represent the selection candidates, ii) updating the TP with data collected on lines from previous breeding cycles before predicting subsequent cycles, and iii) enacting algorithms proposed to identify optimal TP subsets.