Cross-generational decline in genomic selection accuracy in Norway spruce (Picea abies (L.) H. Karst) and strategies to mitigate it

Full abstract

Read the full abstract

Abstract Genomic selection (GS) could be used to reduce the long cycle time for tree breeding. This assumes that marker-based predictions are sufficiently accurate without phenotypes. We evaluated GS across two linked generations of Norway spruce ( Picea abies (L.) H. Karst), the first consisting of 954 plus-trees (G0) and the second of 956 progeny trees representing 34 full-sib families (G1), using 16 clonal field trials across mid- and southern Sweden. Both generations were measured for height and genotyped using a 50 K SNP chip array. Cross-validations within and across generations were compared. GS efficiency was evaluated using global prediction accuracy and within-family predictive ability, using GBLUP with phenotypes for independent validation. We used computer simulations to emulate the experimental data and repeat the same analysis under different assumptions of effective population size. Additional simulations were performed to investigate the cross-generation GS accuracy in future generations. Simulations assuming a historical effective population size of 1000 or 5000, together with experimental results, indicate that prediction accuracy decreased by 49–76% for global prediction and by 15–65% for within-family prediction when G1 was predicted from G0, compared with cross-validation within the G1 generation. Increasing the relatedness at the expense of training set size increased global accuracy but decreased within-family accuracy. Simulations of advanced generations showed that training on multiple generations increases GS accuracy, both for global and within-family prediction. Access to multiple generations for training and/or a higher density of markers may be recommended to increase accuracy for cross-generation and within-family GS in conifers.