Fusarium Head Blight (FHB) is one of the major threatening diseases for wheat in the United States. Genetic progress through breeding has been observed but the phenotypic evaluation of FHB in the field remains tedious. Genomic studies suggest a rather complex genetic architecture with many genes with minor effects involved. For almost 20 years, the Uniform Regional Scab Nursery evaluated breeding lines, mainly from Montana, South Dakota, North Dakota State University, as well as the University of Minnesota.  FHB resistance and other related traits were evaluated in 5 locations and for 30 lines each year on average with 9 locations in total. The experimental design was highly unbalanced since almost all lines are evaluated for one year (across all locations for that year). A subset of these lines has recently been genotyped by two genotypic arrays of 90k and 3k markers. Genomic prediction is a statistical method relying on all available genetic markers to predict the trait of interest and we implemented it to assess its efficiency in predicting FHB susceptibility.

Filters were applied to discard lines and SNPs with too much missing data, imputation was done with Beagle and the minor allele frequency threshold was 5%, leading to 288 lines genotyped for 53k SNPs. Traits were incidence, severity, disease, visual scabby kernel, DON concentration, and heading date, and were analyzed in a linear mixed model.

Genomic prediction was fitted by cross-validation and predictive ability was measured as the Pearson correlation between predicted and observed value in the validation set. The best two methods among the seven tested were rrBLUP (ridge-regression Best Linear Unbiased Predictor) and RKHS (Reproducing Kernel Hilbert Space) across all traits, with an average predictive ability of 0.61 and 0.62, respectively. Highest accuracy was obtained for VSK with a predictive ability of 0.75. Then, we compared the impact of marker density on the predictive ability. We found that predictive ability plateaued with 500 to 3k SNPs, depending on the trait. Finally, we used successively each breeding population (from the different breeding programs) as the validation set, to test the impact of decreasing genetic relatedness. In almost all cases, predictive ability decreased when predicted within a specific breeding program, but various patterns were observed.

These results demonstrate a role for genomic prediction in improving Fusarium head blight resistance traits and the density of markers needed is small, thus allowing use of lower cost genotypic arrays.