Evaluation of genomic and phenomic prediction for application in apple breeding
Apple breeding schemes can be improved by using genomic prediction models to forecast the performance of breeding material. These models predictive ability depends on factors like trait genetic architecture, training set size, relatedness of the selected material to the training set, and the validation method used. Alternative genotyping methods such as RADseq and complementary data from near-infrared spectroscopy could help improve the cost-effectiveness of genomic selection. However, the impact of these factors and alternative approaches on predictive ability beyond experimental populations still need to be investigated. In this study, we evaluated 137 prediction scenarios varying the described factors and alternative approaches, offering recommendations for implementing genomic selection in apple breeding. Our results show that extending the training set with germplasm related to the predicted breeding material can improve average predictive ability across eleven studied traits by up to 0.08. The study emphasizes the usefulness of leave-one-family-out validation, reflecting the application of genomic selection to a new family, although it reduced average predictive ability across traits by up to 0.24 compared to cross-validation. Similar average predictive abilities across traits indicate that imputed RADseq data could be a suitable genotyping alternative to SNP array datasets. The best-performing scenario using near-infrared spectroscopy data for phenomic prediction showed a 0.35 decrease in average predictive ability across traits compared to conventional genomic prediction, suggesting that the tested phenomic selection approach is impractical. These findings offer valuable guidance for applying genomic selection in apple breeding, ultimately leading to the development of breeding material with improved quality.