|RUTKOSKI, JESSICA - Cornell University
|BENSON, JARED - Cornell University
|JIA, YI - Cornell University
|SORRELLS, MARK - Cornell University
Submitted to: The Plant Genome
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/19/2012
Publication Date: 7/15/2012
Citation: Rutkoski, J., Benson, J., Jia, Y., Brown Guedira, G.L., Jannink, J., Sorrells, M. 2012. Evaluation of genomic prediction methods for fusarium head blight resistance in wheat. The Plant Genome. 5:51-61.
Interpretive Summary: Fusarium head blight (FHB) resistance is quantitative and difficult to evaluate. Genomic selection (GS) methods include a number of statistical methods to analyze phenotypes and high density marker data to create models that predict a future phenotype from marker data. Genomic selection (GS) could accelerate FHB resistance breeding. FHB has been evaluated for a number of years in US cooperative nurseries, and this data could be used to train models to help breeding programs improve resistance. We tested a number of GS statistical methods for this purpose. We also tested models that included inexpensive traits to predict deoxynivalenol levels, which are expensive to phenotype. In some models, we included information about previously detected QTL. GS methods gave consistently more accurate predictions than did classical multiple linear regression. Some GS methods were often more accurate than others and are to be recommended. This study indicates that cooperative FHB nursery data can be useful for GS, and prior information about correlated traits and QTL could be used to improve accuracies in some cases.
Technical Abstract: Fusarium head blight (FHB) resistance is quantitative and difficult to evaluate. Genomic selection (GS) could accelerate FHB resistance breeding. We used US cooperative FHB wheat nursery data to evaluate GS models for several FHB resistance traits including deoxynivalenol (DON) levels. For all traits we compared the models: ridge regression (RR), Bayesian LASSO (BL), reproducing kernel Hilbert spaces (RKHS) regression, random forest (RF) regression, and multiple linear regression (fixed effects; MLR). For DON, we evaluated additional prediction methods including bivariate RR models, phenotypes for correlated traits, and RF regression models combining markers and correlated phenotypes as predictors. Additionally, for all traits, we compared different marker sets including: genome-wide markers, FHB quantitative trait loci (QTL) targeted markers, and both sets combined. GS accuracies were always higher than MLR accuracies, RF and RKHS regression were often the most accurate methods, and for DON marker + trait RF regression was more accurate than all other methods. For all traits except DON, using QTL targeted markers alone led to lower accuracies than using genome-wide markers. This study indicates that cooperative FHB nursery data can be useful for GS, and prior information about correlated traits and QTL could be utilized to improve accuracies in some cases.