The study, published in Theoretical and Applied Genetics, was led by Mallikarjuna Rao Kovi of the Norwegian University of Life Sciences together with colleagues at the breeding company Graminor AS. The team evaluated 889 full-sib families of timothy (Phleum pratense L.), a species that accounts for roughly 70 to 80 percent of sown grasslands in Norway and underpins livestock production across the Nordic region. Timothy is an allohexaploid, carrying six sets of chromosomes (2n = 6x = 42) and an estimated genome size of about 4.25 gigabases, making genomic analysis far more complicated than in diploid cereal crops. Allelic dosage at any locus can range from zero to six copies, homoeologous recombination between subgenomes scrambles linkage patterns, and the outcrossing mating system of the species keeps linkage disequilibrium short. These are precisely the conditions under which genomic prediction has been hardest to implement.

The experimental foundation of the work is unusually deep in time. The full-sib families originated from biparental crosses among 49 cultivars and breeding populations, including historically important Norwegian materials such as KGB, Grindstad, Motim and Carola. They were field-tested over a decade, from 2002 to 2012, at two contrasting locations in Southern Norway: Arneberg, a warmer continental inland site at 180 meters elevation, and Løken, a cooler mountain environment at 520 meters. The researchers measured six yield traits, including dry matter yield per cut and total dry matter yield summed across three harvest years, and six quality traits assessed by near-infrared reflectance spectroscopy, covering crude protein, acid detergent fiber, neutral detergent fiber, in vitro digestibility and the Nordic AAT/PBV protein fractions relevant to ruminant nutrition.

Genotyping relied on genotyping-by-sequencing, in which DNA was digested with the ApeKI restriction enzyme, ligated to barcoded adapters and sequenced on an Illumina HiSeq platform. After stringent quality control of an initial 60,635 biallelic SNPs, the final panel comprised 30,698 polymorphic markers across the 889 families. Because diploidized marker encoding has been shown to capture most additive variation in polyploids, the researchers collapsed the six possible dosage classes into a three-level scheme of minus one, zero and one, using the polyRAD algorithm to estimate dosage before diploidization. Population structure analysis revealed only weak relatedness among the families, with the first three principal components explaining less than 1.5 percent of marker variance combined, a feature that would prove consequential later.

At the heart of the study is a systematic comparison of nine prediction models spanning the full methodological spectrum: genomic best linear unbiased prediction implemented through ridge regression, LASSO, Elastic Net, Bayesian ridge, support vector regression, Random Forest, XGBoost, LightGBM and a multi-layer perceptron neural network. Marker information was compressed into principal components before model fitting, retaining up to 150 components for linear models and 100 for machine learning approaches, a strategy that balances information retention against overfitting risk. Within the training population, fivefold cross-validation produced encouraging results, with a mean accuracy of r = 0.62 across all 30 trait-dataset combinations. Random Forest and support vector regression consistently outperformed GBLUP, sometimes by 5 to 15 percent for quality traits, consistent with the idea that nonlinear models can capture epistatic interactions that are particularly relevant in polyploid genomes.

The most striking results came when the researchers turned from cross-validation to forward validation. They established 213 entirely independent full-sib families in 2015, evaluated at both locations, that shared zero overlap with the training set. When models trained on the earlier data were applied to these genuinely novel families, mean accuracy collapsed to r = 0.16, a generalization gap of 0.46, or roughly a 74 percent reduction relative to cross-validation estimates. Only 16 of the 30 trait-dataset combinations reached statistical significance. The best-performing combinations still offered usable signal: second-cut dry matter yield at Arneberg predicted with support vector regression reached r = 0.52, total dry matter yield with GBLUP achieved r = 0.25, and the rumen protein balance trait PBV reached r = 0.42 in the first cut. But other traits, such as third-cut yield and second-year first-cut yield at the mountain site, fell to near zero or even negative correlations.

The authors attribute the gap to several reinforcing factors. Full-sib family structure means that when related families appear in both training and testing folds during cross-validation, models can exploit family-level associations rather than genuine marker-trait linkage, inflating apparent accuracy. The diploidized encoding of a hexaploid genome may also capture population-specific linkage patterns that fail to transfer across breeding cycles, and year-to-year environmental variation between the training and validation trials adds noise. Notably, simpler models like GBLUP and support vector regression tended to generalize better than complex ensembles, a classic bias-variance tradeoff in which the extra flexibility of Random Forest and gradient boosting buys within-sample fit at the cost of out-of-sample robustness. The finding that the best cross-validation model was often not the best forward-validation model is a caution that should resonate well beyond forage grass breeding.

The study also quantified the genomic architecture underlying these predictions. Genomic heritabilities, estimated by GREML from a VanRaden genomic relationship matrix across environments, were moderate for yield traits, ranging from 0.37 for total dry matter yield to 0.55 for third-cut yield. Quality traits told a different story: heritabilities ranged from 0.14 for rumen protein balance down to effectively zero for intestinal amino acid content, explaining the weak forward validation observed for many quality measures and the fact that genetic correlations among quality traits could not be estimated reliably. Among yield traits, correlations were clearer, with total yield almost perfectly correlated with first-cut components and a weakly negative correlation between second-cut yields in consecutive years, highlighting year-specific genetic control of regrowth.

Practical design parameters emerged clearly from the data. Prediction accuracy plateaued at approximately 15,000 SNPs, meaning a cost-effective panel of 15,000 to 20,000 markers would capture 90 to 95 percent of maximum achievable accuracy. Training population size followed a diminishing-returns curve, with 250 to 300 families capturing most available accuracy. Multi-trait models delivered modest but consistent gains of 3 to 5 percent, and genotype-by-environment interaction models that explicitly incorporated location as a covariate alongside marker-by-location interaction terms achieved the highest within-training accuracy of the study, r = 0.71 with Random Forest, underscoring the importance of environment-specific modeling in a species known to be sensitive to photoperiod and temperature.

The work also moved beyond diagnosis into application. Using genomic estimated breeding values from the training population, a multi-trait selection index weighted toward total dry matter yield, digestibility and crude protein identified the top 30 families for crossing, most of them derived from KGB backgrounds. The highest-scoring recommended cross, Geno_363 by Geno_368, balanced a high-yielding KGB pedigree against Grindstad-derived parentage, combining genetic merit with diversity to exploit potential heterosis while penalizing within-family matings to limit inbreeding. For the most predictable trait, families ranked in the top 20 by genomic estimated breeding value showed a mean phenotypic advantage of 107 kilograms of dry matter per hectare over the validation population mean.

The researchers caution that several limitations temper these benchmarks. The diploidized encoding discards allelic dosage information, the principal component compression retains only about 18 to 25 percent of marker variance, and with each family represented by essentially a single field record per environment, genomic and residual variances are only weakly separable within any one location. Still, the practical message is constructive. Even with realistic forward-validation accuracies between 0.20 and 0.50 for the most tractable traits, genomic selection can accelerate genetic gain by enabling early screening of unphenotyped seedlings and shortening the evaluation cycles needed before cultivar release. The team recommends diverse training populations of at least 300 families, rigorous forward validation before committing to a model, cost-effective marker panels, robust baseline models like GBLUP and support vector regression, and a two-stage scheme that reserves field trials for the top genomic candidates. For a species central to Northern European agriculture and increasingly relevant as climate pressure reshapes forage systems, this study sets the benchmark against which future genomic breeding efforts in polyploid grasses will be measured.