min.lin@nmbu.no

Yellow Rust (YR) in wheat, caused by Puccinia striiformis f. sp. tritici, can significantly affect global wheat production, especially in cool and humid regions. Genomic-assisted breeding, utilizing genetic markers, is effective in selecting breeding lines with high breeding values and enhanced disease resistance. In this study, 300 Nordic spring wheat lines were genotyped using the wheat 25K SNP chip, and their resistance to YR was tested across 17 trials conducted on four continents over seven years.

The objective of this study was to apply genomic prediction for yellow rust resistance and compare the prediction ability (PA) of different models. Three genomic prediction models were assessed: Model 1 utilized only significant markers from the Genome-Wide Association Study (GWAS) as genomic information (Gs), alongside environments (E), and the genomic by environment interaction (Gs × E) to fit the Genomic Best Linear Unbiased Prediction (GBLUP) model. Model 2 employed the entire genomic data (G) and fitted the GBLUP model with E and G × E. Model 3 was a deep learning model that utilized complete genome data and environments.

Cross-validation (CV) was performed by randomly selecting 20% of observations as the test set and using the remaining data as the training set. This procedure was repeated five times. Through CV, the PA for each environment varied, ranging from 0.32 to 0.84 for Model 1, 0.41 to 0.87 for Model 2, and 0.45 to 0.97 for Model 3.

Additionally, we predicted four trials from 2021 using data from thirteen trials conducted from 2015 to 2020 for training. The PAs with Model 1 were 0.54, 0.47, 0.64, and 0.65; with Model 2 were 0.63, 0.59, 0.70, and 0.73; and with Model 3 were 0.61, 0.56, 0.72, and 0.73, respectively. These results indicate that a reasonable PA can be achieved even when only significant markers are utilized as genomic information for prediction.

However, the inclusion of complete genomic information in Model 2 could further enhance the PA up to 70%. Furthermore, although there were observable differences in the PA of individual trials between Model 2 and Model 3, these models yielded comparable results.