Evaluation of Genomic Prediction Methods for Fusarium Head Blight Resistance in Wheat

Breeding for resistance to Fusarium Head Blight (FHB) is slow and costly because it is quantitatively inherited and difficult to evaluate accurately. A new marker assisted breeding method, genomic selection (GS), which is suited for quantitative traits, could potentially accelerate breeding for FHB resistance. GS involves predicting breeding values based on genome-wide markers using a model trained with phenotypic and genotypic data. We used data from the cooperative Fusarium head blight (FHB) nurseries across the United States to evaluate prediction accuracies for FHB resistance traits: severity (SEV), incidence (INC), Fusarium damaged kernels (FDK), and incidence/severity/kernel quality index (ISK), as well as deoxynivalenol levels (DON) and days to heading (HD). For all traits we compared prediction accuracies of four different GS models: Ridge-Regression (RR), Bayesian-Lasso (BL), Reproduction Kernel Hilbert Spaces Regression (RKHS), Random Forest Regression (RF), and one multiple linear regression model (MLR). In addition, we compared prediction accuracies using three different marker sets: genome-wide markers, FHB QTL targeted markers only, and both sets combined. We found that GS accuracies were always higher than MLR accuracies, and except for DON, using QTL targeted markers alone always led to lower accuracies. For DON, a low heritability and expensive to measure trait, we also evaluated prediction accuracies achieved from using phenotypes for correlated traits: SEV, INC, FDK, and ISK, as well as a RF model combining markers and ISK as predictors. The marker-only RF model predicted DON as accurately as the correlated traits, while the marker + trait RF model was significantly more accurate than any other prediction method tested. Our results showed that we can expect genetic gain from implementing GS for FHB resistance in this germplasm, and for DON, we can expect equal genetic gain per cycle and greater genetic gain per unit time with GS compared to selection based on correlated traits.