Genomic Selection in Tree Breeding: Prediction Accuracy in a White Spruce Diallel Mating Design

The ability to predict the genetic value of trees at an early age is of interest to tree breeders coping with long selection cycles in boreal conifers. Mature phenotypes, such as wood traits, can only be assessed after many years of testing and marker-based selection tools are thought to be a cost-effective alternative to accelerated breeding. A genomic selection (GS) analysis was carried out in a population of 303 clones derived from 75 full-sib families produced using a disconnected diallel mating design, with 20 parents divided into 4 groups. The effectiveness of GS models was evaluated by the ability to predict phenotypes and breeding values using two types of population subset. First, a 10-fold cross-validation was performed in which full-sib families are included in each training set. Second, three breeding groups were used for training and GS models were validated in the remaining group. For the first validation scheme, marker‑predicted breeding values were well correlated with “true” breeding values, as estimated using available pedigree information for all 303 clones. Correlations for cross‑validated pedigree models were somewhat stronger, but prediction accuracies obtained with markers were encouraging. The ability to predict phenotypes was weak to moderate and carried large errors. Using trees from an unrelated diallel group for validation led to a drop in accuracy close to 0 as expected from earlier studies on white spruce. Results underline the importance of a shared relationship between training and validation sets of GS models to predict breeding value or phenotypes.