Cross-generational genomic prediction of Norway spruce (Picea abies) wood properties: an evaluation using independent validation.

Background: The evaluation of genomic selection (GS) efficiency in forestry has primarily relied on cross-validation schemes that split the same population within a single generation for both training and validation. While useful, this approach may not be reliable for multigenerational breeding. To our knowledge, this is the first study to assess genomic prediction in Norway spruce using a large dataset spanning two generations in two environments. We trained pedigree-based (ABLUP) and marker-based (GBLUP) prediction models under three approaches: forward prediction, backward prediction, and across-environment prediction. The models were evaluated for ring-width, solid-wood and tracheid characteristics, using ~ 6,000 phenotyped and ~ 2,500 genotyped individual. Predictive ability (PA) and prediction accuracy (ACC) were estimated using an independent validation method, ensuring no individuals were shared between training and validation datasets. To assess the trade-off between comprehensive radial history and practical direct methods, we compared GBLUP models trained with cumulative area-weighted density (AWE-GBLUP) and single annual-ring density (SAD-GBLUP) from mother plus-trees. These models were validated using early and mature-stage progeny density measurements across two trials.

Results: Despite the smaller number of individuals used in the GBLUP models, both PA and ACC were generally comparable to those of the ABLUP model, particularly for cross-environment predictions. Overall, forward and backward predictions were significantly higher for density-related and tracheid properties, suggesting that across-generation predictions are feasible for wood properties but may be challenging for growth and low-heritability traits. Notably, SAD-GBLUP provided comparable prediction accuracies to AWE-GBLUP, supporting the use of more practical and cost-effective phenotyping methods in operational breeding programs.

Conclusions: Our findings highlight the need for context-specific models to improve the accuracy and reliability of genomic prediction in forest tree breeding. Future efforts might aim to expand training populations, incorporate non-additive genetic effects, and validate model performance across cambial ages while accounting for climatic variability during the corresponding growth years. Overall, this study offers a valuable foundation for implementing GS in Norway spruce breeding programs.