The research team, led by Jared Lile with Jennifer Lachowiec, Jason D. Fiedler of the USDA Agricultural Research Service, and Jason P. Cook, mined twelve years of historical quality data from the Montana State University Spring Wheat Breeding Program, spanning 2010 through 2022. In total, 842 lines were phenotyped, including 24 check varieties and 818 elite experimental lines of hard red spring wheat, a market class that accounted for roughly a quarter of total United States wheat production in 2024. Hard red spring wheat is prized for its high grain protein content and gluten strength, and its market price depends directly on how well the flour performs in the mill, the mixer, and the oven.

The twelve traits under scrutiny spanned the full quality spectrum, from milling characteristics such as flour yield and ash content, through dough properties like mix time, mix water absorption, and mix tolerance, to baking outcomes including bake mixing time, bake water absorption, and loaf volume. Measurements followed American Association of Cereal Chemists protocols, using near-infrared reflectance instruments for protein determination, a Brabender Quadrumat Sr. mill for flour yield, a Mixograph for dough rheology, and standardized bake tests for loaf volume. Single kernel hardness was captured separately with a single-kernel characterization system. This depth of phenotyping, sustained over more than a decade, is what gave the study its statistical power.

On the genetic side, the team used genotyping-by-sequencing to profile all 842 lines, ultimately retaining 21,958 single nucleotide polymorphisms after filtering and imputation against the Chinese Spring wheat reference genome. Broad-sense heritability, calculated with the Cullis method suited to unbalanced datasets, proved high for nearly every trait: all but one exceeded 0.60, with only wheat ash content falling to 0.39. That combination of high heritability and polygenic architecture, meaning many genes of small effect, makes end-use quality traits ideal candidates for genomic selection, which models the entire genome simultaneously rather than tracking a handful of major loci as marker-assisted selection does.

The researchers compared two broad analytical frameworks. The first was a conventional two-step approach: best linear unbiased estimators were computed for each line across all environments, and those summary values then served as response variables for seven genomic prediction models, including ridge regression BLUP, Bayesian ridge regression, Bayesian LASSO, BayesA, BayesB, BayesC, and reproducing kernel Hilbert space models. In five-fold cross-validation repeated over 100 cycles, mean predictive abilities ranged from 0.4 to 0.7, with mix tolerance and flour yield performing best and the low-heritability wheat ash content performing worst. Strikingly, the choice of model mattered far less than the trait itself; within any given trait, the seven models were nearly indistinguishable.

The second framework was a single-step reaction norm model, an extension of the GBLUP approach that feeds raw phenotypic observations and marker data into one unified model. Its key innovation is an explicit genotype-by-environment interaction term, built from covariance matrices that combine the genomic relationship matrix with environmental incidence matrices. This allows the model to predict how a specific line will perform in a specific environment, mimicking a scenario in which a breeder phenotypes only a few check varieties in a trial and predicts the rest. In cross-validation, the median predictive ability within location-year combinations tracked the heritability of each trait almost exactly, with bake time, flour yield, mix time, and single kernel hardness approaching 0.8.

The team also tested whether incorporating known major genes as fixed effects could sharpen predictions. Wheat end-use quality is strongly influenced by seed protein families, including glutenins, gliadins, puroindolines, purothionins, avenin-like proteins, purinins, triticins, and amylase-trypsin inhibitors. Markers within one megabase of these genes were pulled out of the random-effect marker matrix and treated as fixed effects, either individually, as principal components of the candidate marker set, or as markers discovered by genome-wide association studies run within each training fold. None of these strategies improved predictive ability, a result the authors attribute to the fact that the genome-wide models were already capturing the effects of these loci implicitly.

One cautionary finding deserves wide attention. When the researchers included GWAS markers identified on the full dataset, spanning both training and validation material, predictive ability jumped by roughly one standard deviation across all traits. That apparent gain was an artifact of data leakage, sometimes called the inside trading effect, in which information from the validation set contaminates model fitting. The demonstration is a timely reminder for the genomics community that benchmarking practices matter: fixed-effect markers must be identified strictly within training data, or reported accuracies will be silently inflated.

The most demanding test was leave-one-year-out validation, in which all experimental lines from an entire year were withheld and predicted from the remaining eleven years of data. Using a predictive ability threshold of 0.3 as the boundary between useful and uninformative predictions, five traits, namely bake time, loaf volume, mix time, mix tolerance, and single kernel hardness, cleared the bar in all thirteen years. Wheat ash content was the only trait that consistently fell short. In the forward prediction of 2022, the most recent year, the reproducing kernel Hilbert space model showed a discernible edge for grain and flour protein content, though a gap in genotyping for the 2021 preliminary yield trial lines likely dampened overall performance in that final year.

The practical implications reach well beyond Montana. Because the two-step and single-step methods answer different questions, one estimating a line’s overall breeding value across environments, the other predicting its phenotype in a specific environment, breeders can deploy them at different stages of the pipeline. By selecting on genomic predictions before expensive quality testing is possible, programs can increase selection intensity for end-use quality and accelerate genetic gain, ensuring that the bread wheat reaching global markets continues to meet the exacting standards that millers, bakers, and consumers demand. As training datasets accumulate and reference genomes improve, the authors expect these approaches to become routine tools for keeping one of the world’s most important staple crops a step ahead.

Subject of Research: Genomic selection models for predicting end-use quality traits in hard red spring wheat breeding

Article Title: Evaluating reaction norm and fixed effect models for genomic selection of end-use quality in wheat

Article References: Lile, J., Lachowiec, J., Fiedler, J. D., & Cook, J. P. (2026). Evaluating reaction norm and fixed effect models for genomic selection of end-use quality in wheat. Theoretical and Applied Genetics, 139(10), Article 293. https://doi.org/10.1007/s00122-026-05394-4

Image Credits: AI Generated

DOI: 10.1007/s00122-026-05394-4

Keywords: genomic selection, wheat, end-use quality, hard red spring wheat, reaction norm model, genotype-by-environment interaction, fixed effects, GWAS, predictive ability, breeding, baking quality, GBLUP