Genomic Selection in Citrus Breeding: Accuracy of Genomic Prediction

Cross breeding of citrus is strongly hindered by their long lifespan, large plant size and extended juvenile phase for seedlings. Genomic selection is an attractive technology to surmount the problems because it enables selection without field-testing. In this study, we evaluated the accuracy of genomic prediction using 106 citrus varieties. Linkage disequilibrium in the varieties extended around 20 Mb. The accuracy of genomic prediction with four linear and two non-linear models was assessed via cross-validation in 17 traits. Correlation coefficient between observed and predicted genetic values was high (> 0.7) in fruit weight, fruit hardiness, ease of peeling, fruit color and fruit flesh color. Linear and Gaussian kernel regression models showed stable accuracy over the traits. To assess the potential of the population for detecting loci controlling traits, we performed genome-wide association study (GWAS). GWAS detected significant associations in all the traits, suggesting the existence of major QTL controlling the traits. However, genomic prediction with the model treating significant SNPs as fixed effects and the remaining SNPs as random effects was less accurate than genomic prediction with the model treating all SNPs as random effects in most traits, suggesting the robustness of genomic prediction based on the whole-genome markers. We assessed the prediction accuracy in the tails of the distribution of fruit size, and found linear kernel was more accurate than Gaussian kernel in the tails. The results suggest that genomic selection is promising in citrus breeding and that linear kernel regression is recommended as a model for genomic prediction.