The research team, led by J. A. Woolliams and Gregor Gorjanc of The Roslin Institute at the University of Edinburgh, together with colleagues at the University of Oxford and Forest Research, analysed 1295 elite spruce trees drawn from 50 full-sib families, along with 73 open-pollinated control trees from a Haida Gwaii provenance. The trees came from two genetically linked field trials established in 1996 and 1997 and replicated at two contrasting sites: Brecon in Wales at 375 metres above sea level, and Kintyre in Scotland at just 60 metres. Nine traits were measured, spanning early growth traits such as height at six years and diameter at eleven years, visual scores for branching and stem straightness, a pilodyn penetration reading used as a proxy for wood density, and, on felled trees, merchantable volume, modulus of rupture, modulus of elasticity and wood density. The wood quality traits were assessed on small clear specimens cut from boards spanning the diameter of each log, stabilised to 12 percent moisture content and tested in three-point flexural bending, with cross-sectional averages weighted by ring area to reflect how European softwood grading actually values timber.

The conceptual heart of the study lies in the distinction between two kinds of genetic information. Pedigree-based evaluation, the industry standard since the work of Sewall Wright in the 1920s, relies on identity by descent: the expected probability that two relatives share copies of the same ancestral gene, derived purely from recorded family relationships. Genomic evaluation, pioneered in livestock by Meuwissen and colleagues in 2001, instead typically uses identity by state, in which thousands of DNA markers scattered across the genome capture the actual sharing of chromosome segments between individuals, exploiting population-wide linkage disequilibrium between markers and the causal variants underlying traits. The team compared three relationship matrices within a single statistical framework: the classical pedigree matrix, a VanRaden genomic matrix built from identity by state, and a third, less common option that uses dense markers in linkage analysis to track the realised inheritance of genome segments from offspring back through the pedigree to its founders, yielding a genomic identity-by-descent matrix.

Genotyping was carried out on a SNP chip containing 12,911 markers, of which 4123 could be assigned to the 12 linkage groups of a recently published Sitka spruce genetic map spanning 21.42 Morgans. Missing genotypes were filled in by multi-locus peeling using the AlphaPeel software, raising the proportion of trees with call rates above 95 percent from 84 to 96 percent, with an estimated residual genotype error probability of 0.022. The pedigree itself was audited against the genomic data, revealing inconsistencies for two parents that required dummy parents to be introduced into the pedigree for four families, a quiet reminder of how often recorded pedigrees in tree breeding diverge from biological reality.

The critical methodological innovation was in how accuracy was judged. Model-based accuracies, calculated from prediction error variances under the fitted mixed model, assume that both the model structure and its estimated parameters are correct. Cross-validation makes no such assumption: the researchers withheld phenotypes from roughly one-twelfth of the elite trees at a time in a 12-fold scheme, re-estimated breeding values without those records, and then correlated the predictions with the withheld phenotypes block by block across the two sites, pooling the results through a meta-analytic framework that accounted for heterogeneity among blocks. This distinction proved decisive. When accuracies were read directly from the fitted models, pedigree and genomic approaches looked almost indistinguishable. But under the more objective cross-validation, genomic information delivered the highest accuracy for all but one of the nine traits, and the pedigree-based model showed a characteristic weakness: without genomic data, two full siblings in the validation set receive identical breeding values, because the only source of information about an individual’s Mendelian sampling, its own phenotype, has been deliberately withheld.

That last point gets to the biological and statistical core of why genomics helps. Within a full-sib family, each offspring inherits a unique random shuffle of its parents’ chromosomes, and this Mendelian sampling term is the engine of within-family genetic variation. Pedigree can only assign siblings the same expected relationship; markers can observe which sibling actually inherited which segments. The team exploited this with a fourth model, IIIA, that separated the between-family component, captured by parental transmitting abilities, from the within-family Mendelian sampling component, estimated directly from the markers. This model produced the best cross-validation accuracies, suggesting that in breeding programmes with large sib families, as is typical in conifers, genomic selection can be made to work efficiently even with a relatively sparse marker panel, because recombination events are rare enough along a chromosome that localising them does not require very high marker density.

The study was also candid about the limitations of its genomic data. Several independent lines of evidence indicated that the 4123 markers failed to capture the full genetic variance. Genomic models yielded lower heritability estimates than the pedigree model for traits with high heritability, with the unaccounted genetic covariance among siblings leaking into the residual variance. More tellingly, the ratio of Mendelian sampling variance to the classical expectation under random mating was estimated at 0.666, implying that the markers captured only about 83 percent of the true genetic variance. For context, livestock breeders routinely work with around 2500 markers per Morgan, roughly ten times the density used here. Yet despite this handicap, the genomic models still matched or beat pedigree accuracy, a result the authors attribute to the rich sib structure of the population and the value of estimating Mendelian sampling directly.

Equally important was what the study did not find. Some researchers have hypothesised that pedigree-based breeding values are systematically biased relative to genomic ones, but across all trait and information combinations, the regression slopes of phenotype on breeding value fell within 1.5 standard errors of one, indicating no evidence of over- or under-estimation in any model. The authors note that theory supports this: with the correct model, unbiased estimates should follow, and both pedigree and genomic approaches remain vulnerable to unmodelled sources of field variation such as maternal effects or environmental gradients. The absence of bias strengthens confidence that the accuracy gains from genomics are genuine rather than artefacts of miscalibration.

The forward-looking implications are where the study becomes genuinely striking. With the current training set of roughly 1341 genotyped and phenotyped individuals, cross-validation accuracies of genomic breeding values ranged from about 0.6 to 0.7 across traits. Projections using established theory suggest that an eightfold increase in the training population would push accuracies to 0.8 or 0.9 for all traits studied, approaching the ceiling imposed by the current marker density. Pedigree-based accuracy, by contrast, is fundamentally capped: an eightfold increase in records per parent would lift validation accuracy only from about 0.5 to 0.667, because predictions for unphenotyped individuals rest entirely on parental averages. In other words, genomics is already competitive at a scale where pedigree has nearly exhausted its potential, and every additional investment in training data and marker density compounds the genomic advantage rather than hitting a wall.

For forestry, the practical consequences could be transformative. Genomic breeding values can be computed for seedlings or even seeds before any field trial, enabling early selection decisions that shave years off breeding cycles, and for traits such as wood strength that require felling a tree to measure properly, genomics offers selection without destruction. The authors argue that realising this potential in conifers will require denser SNP panels to capture the full genetic variance in less structured populations, and training sets numbering in the tens of thousands rather than thousands. But the direction of travel is now clear. Sitka spruce, with a genome larger than ten billion base pairs and a breeding history written more in field notebooks than in sequence data, has just demonstrated that the genomic era of tree breeding is no longer a promise for the future. It is a working tool, ready to be scaled.

Subject of Research: Genomic versus pedigree-based estimation of breeding values in Sitka spruce

Article Title: Genomic models for accurate estimation of breeding values in Sitka spruce (Picea sitchensis Bong. Carr)

Article References: Genomic models for accurate estimation of breeding values in Sitka spruce (Picea sitchensis Bong. Carr). (n.d.). https://doi.org/10.1038/s41437-026-00894-x

Image Credits: AI Generated

DOI: 10.1038/s41437-026-00894-x

Keywords: genomic selection, Sitka spruce, breeding values, identity by descent, identity by state, conifer breeding, wood quality traits, heritability, cross-validation, Mendelian sampling, SNP markers, forest genetics