Abstract
Genomic selection (GS) has the potential to improve selection efficiency and shorten the breeding cycle in fruit tree breeding. In this study, we evaluated the effect of prediction methods, marker density and the training population (TP) size on pear GS for improving its performance and reducing cost. We evaluated GS under two scenarios: (1) five-fold cross-validation in an interspecific pear family; (2) independent validation. Based on the cross-validation scheme, the prediction accuracy (PA) of eight fruit traits varied between 0.33 (fruit core vertical diameter) and 0.65 (stone cell content). Except for single fruit weight, a slightly better prediction accuracy (PA) was observed for the five parametrical methods compared with the two non-parametrical methods. In our TP of 310 individuals, 2 000 single nucleotide polymorphism (SNP) markers were sufficient to make reasonably accurate predictions. PAs for different traits increased by 18.21%–46.98% when the TP size increased from 50 to 100, but the increment was smaller (-4.13%–33.91%) when the TP size increased from 200 to 250. For independent validation, the PAs ranged from 0.11 to 0.45 using rrBLUP method. In summary, our results showed that the TP size and SNP numbers had a greater impact on the PA than prediction methods. Furthermore, relatedness among the training and validation sets, and the complexity of traits should be considered when designing a TP to predict the test panel.
