Genomic Selection Speeds Breeding Wheat Resistant to Two Rust Diseases

Wheat breeders in the Czech Republic have shown that a genome-wide statistical approach can predict resistance to two of the world’s most damaging wheat diseases, leaf rust and stem rust, well enough to pick promising breeding lines before they ever go through years of field testing. The study, published in the Journal of Agriculture and Food Research, combined visual disease scores with direct molecular measurements of how much fungus had colonised each plant, then used that data to train models that estimate the breeding value of young wheat lines from their DNA alone.

The research team, led by Leona Leišová-Svobodová, worked with 186 wheat varieties and lines as a training population and 103 advanced breeding lines as the population to be predicted. All material came from the SELGEN breeding programme. Field experiments ran at two sites on opposite edges of Prague, Stupice and Prague-Ruzyně, over a period stretching from 2003 to 2025, with spring wheat accessions additionally evaluated in 2022 and thus assessed across four years. At each location, two separate field trials were established, one for stem rust caused by Puccinia graminis f. sp. tritici and one for leaf rust caused by Puccinia triticina, each with two randomised replications and spatially separated to prevent mixed infection.

Disease was deliberately encouraged rather than avoided. Spreader rows of susceptible varieties, Michigan Amber, Baletka and Victo, promoted uniform infection, and plants were artificially inoculated at roughly the growth stage BBCH 29 to 37 with urediniospore mixtures representing races with defined virulence profiles, collected from infected wheat leaves across the Czech Republic in the previous season and multiplied on susceptible seedlings in the greenhouse. Resistance was scored on the standard Czech 1-to-9 scale, where 1 means highly susceptible and 9 highly resistant. In parallel, the team quantified fungal biomass directly: five plants per plot were sampled, DNA was extracted with an optimised CTAB protocol, and real-time PCR with pathogen-specific primers and TaqMan probes measured the ratio of fungal to wheat DNA. This produced four traits per genotype: visual scores and pathogen content for each rust species.

The phenotypic analysis revealed a striking asymmetry between the two diseases. Broad-sense heritability was high for stem rust traits, 0.95 for the visual evaluation and 0.79 for pathogen content, but lower for leaf rust, at 0.80 and 0.58 respectively. The correlation between visual score and pathogen content was 0.70 for stem rust but only 0.53 for leaf rust, and pathogen content did not always fully correspond to symptom severity, particularly for leaf rust. Most tellingly, resistance to the two rusts was barely related to each other at all: correlations between stem-rust and leaf-rust traits were just 0.14 for visual evaluations and −0.17 for pathogen content. That near-independence reflects largely separate genetic control of resistance to the two Puccinia species, and it is precisely what makes combining both resistances in one variety so difficult with conventional selection.

Genotyping was carried out on the DArTseq platform. The training population was profiled with a high-density panel of 103,307 SNPs, while a cheaper panel of 55,612 SNPs was used for the breeding lines. After quality filtering, which removed markers with more than 10 percent missing data or a minor allele frequency below 5 percent, the training dataset retained 185 accessions and 13,275 markers, and a combined dataset of 288 genotypes and 10,092 shared markers was used for prediction. Population structure analysis showed only weak differentiation among the training accessions, with substantial admixture, which the authors argue supports the set’s suitability as a training population for genomic prediction.

Genomic selection, first proposed as an extension of marker-assisted breeding, works in two steps. A training population with both phenotypes and genome-wide markers is used to calibrate a statistical model linking marker variation to trait variation; the model then predicts genomic estimated breeding values, or GEBVs, for selection candidates that have not yet been extensively phenotyped. Because rust resistance in this material was previously shown to depend on many loci with small effects rather than a few major genes, the researchers compared Bayesian models that allow marker effects to differ in size. Six single-trait models were tested, and Bayes A emerged as the best fit by deviance information criterion and residual variance.

Prediction accuracy, measured as the correlation between predicted breeding values and phenotypic BLUPs across 100 repeated cross-validation runs, was strong for stem rust and modest for leaf rust. The Bayes A model reached 0.79 for visual stem rust evaluation and 0.75 for stem rust pathogen content, but only 0.42 and 0.38 for the corresponding leaf rust traits. The multi-trait model, a Bayesian ridge regression that analyses all four traits simultaneously with an unstructured covariance matrix, changed that picture: leaf rust prediction accuracy rose to 0.47 for visual evaluation and 0.46 for pathogen content, while stem rust accuracy stayed essentially unchanged at 0.75 and 0.72. The improvement for leaf rust traits illustrates the core promise of multi-trait genomic selection, that information from correlated traits can boost predictions for harder-to-predict ones, though here the benefit flowed in only one direction.

Correlations between GEBVs from the single-trait Bayes A and multi-trait BRR approaches were extremely high for all four variables, between 0.984 and 0.995, indicating that both models ranked genotypes in nearly the same order. Selection of candidate lines then went beyond raw predictions. The team calculated a posterior expected loss for each breeding line using the MPS R package, and combined this with a genetic similarity matrix based on simple matching coefficients. Thresholds were set at the 35th percentile of expected loss and the 65th percentile of genetic similarity, deliberately flagging roughly 20 lines that combined favourable predictions with low genetic redundancy, so that genetic diversity in the breeding programme would not be squeezed out by selection.

The payoff was quantified. Relative to non-selected lines, the chosen candidates showed an estimated improvement in average GEBVs of 16.1 percent for visual stem rust evaluation, 4.9 percent for visual leaf rust evaluation, 3.1 percent for stem rust pathogen content and 0.5 percent for leaf rust pathogen content. The authors note that prediction accuracy in genomic selection depends on trait heritability, marker density, relatedness between training and breeding populations and training population size, which is why accuracies vary widely across published studies; comparable figures elsewhere range from roughly 0.12 to 0.71 for leaf rust and 0.27 to 0.65 for stem rust depending on method and material.

The practical message is that genomic estimated breeding values can identify promising parental and breeding lines early in the cycle, reducing dependence on repeated, labour-intensive and subjective field phenotyping, with particularly high reliability for stem rust resistance. The combination of genomic prediction, a posterior expected loss function and genetic similarity filtering offers breeders a framework for improving resistance to both rust diseases at once while keeping the genetic base of their material broad. The roughly 20 selected lines now move forward for further testing and potential use in subsequent breeding cycles, a modest number of candidates carrying predicted gains against two pathogens that have shadowed wheat cultivation for millennia.

Subject of Research: Genomic selection for combined leaf and stem rust resistance in wheat breeding

Article Title: Multi-trait genomic selection for combined leaf and stem rust resistance in wheat breeding

Article References: Leišová-Svobodová, L., Hanzalová, A., Veškrna, O., Jungová, M., Sovová, T., & Svoboda, P. (2026). Multi-trait genomic selection for combined leaf and stem rust resistance in wheat breeding. Journal of Agriculture and Food Research, 32, Article 103381. https://doi.org/10.1016/j.jafr.2026.103381

Image Credits: AI Generated

DOI: 10.1016/j.jafr.2026.103381

Keywords: wheat, genomic selection, leaf rust, stem rust, Puccinia triticina, Puccinia graminis, GEBV, Bayes A, Bayesian ridge regression, multi-trait model, DArTseq, plant breeding