Genomic selection in timothy (Phleum pratense L.): a comprehensive evaluation of prediction models, multi-trait strategies, and forward validation across Norwegian environments
Evaluation study shows substantial prediction accuracy drops during forward validation in hexaploid timothy, indicating cross-validation overestimates real-world forage breeding gains.
Key Points
To evaluate genomic prediction models, multi-trait selection strategies, and forward validation performance for yield and quality traits in hexaploid timothy across multiple environments.
Evaluated 889 full-sib (FS2) families derived from 49 cultivars/populations across three harvest years at a highland and a lowland site in Southern Norway.
Genotyped families using 30,698 single nucleotide polymorphism (SNP) markers obtained via genotyping-by-sequencing (GBS).
Benchmarked nine genomic prediction models across six yield and six quality traits using within-training cross-validation and forward validation on 213 independent FS2-families.
Within-training cross-validation yielded moderate to high accuracy (mean r = 0.62; mean r = 0.71 for family-by-environment Random Forest models), but forward validation accuracy dropped substantially to a mean of r = 0.16 (p < 0.05 in 16 of 30 trait-dataset evaluations).
Genomic heritabilities ranged from near zero for quality traits to 0.55 for yield traits, while multi-trait models improved prediction accuracy by 3–5% over single-trait models and marker density plateaued at approximately 15,000 SNPs.