Within-family genomic selection in strawberry: Optimization of marker density, trial design, and training set composition

Genomic selection is a widely used quantitative method of determining the genetic value of an individual from genomic information and phenotypic data. In this study, we used a large, multi-year training population of 3248 individuals from the University of Florida strawberry (Fragaria x ananassa Duchesne) breeding program. We coupled this training population with a test population of 1460 individuals derived from 20 biparental families. Using these two populations, we tested different genomic selection methods of predicting each family separately for the purpose of within-family selection of seedlings for multiple yield-related traits in strawberry. The methodology we considered were comprised of 11 different marker densities, 10 different training set sizes, four different training set composition techniques, and one to five clonal replications for each individual in the training population. We demonstrated that prediction accuracy varied among the 20 biparental families from 0.05 to 0.63 for the three traits investigated. We also showed that a medium-density genotyping strategy (1500-1650 single nucleotide polymorphisms) could be 95%-97% as effective as a high-density genotyping platform and that imputation to the more dense platform always improved accuracy. Training set composition techniques had no discernible effect on prediction accuracy. However, increasing training set size improved prediction accuracy, and accuracy did not plateau even when training sets exceeded 3000 individuals. Finally, we showed that the number of clonal replicates in field trials could be reduced by 80% without any negative effects on genomic selection accuracy.