Comparison of Genomic Selection Models to Predict Flowering Time and Spike Grain Number in Two Hexaploid Wheat Doubled Haploid Populations

Genomic selection (GS) is becoming an important selection tool in crop breeding. In this study, we compared the ability of different GS models to predict time to young microspore (TYM), a flowering time-related trait, spike grain number under control conditions (SGNC) and spike grain number under osmotic stress conditions (SGNO) in two wheat biparental doubled haploid populations with unrelated parents. Prediction accuracies were compared using BayesB, Bayesian least absolute shrinkage and selection operator (Bayesian LASSO / BL), ridge regression best linear unbiased prediction (RR-BLUP), partial least square regression (PLS), and sparse partial least square regression (SPLS) models. Prediction accuracy was tested with 10-fold cross-validation within a population and with independent validation in which marker effects from one population were used to predict traits in the other population. High prediction accuracies were obtained for TYM (0.51–0.84), whereas moderate to low accuracies were observed for SGNC (0.10–0.42) and SGNO (0.27–0.46) using cross-validation. Prediction accuracies based on independent validation are generally lower than those based on cross-validation. BayesB and SPLS outperformed all other models in predicting TYM with both cross-validation and independent validation. Although the accuracies of all models are similar in predicting SGNC and SGNO with cross-validation, BayesB and SPLS had the highest accuracy in predicting SGNC with independent validation. In independent validation, accuracies of all the models increased by using only the QTL-linked markers. Results from this study indicate that BayesB and SPLS capture the linkage disequilibrium between markers and traits effectively leading to higher accuracies. Excluding markers from QTL studies reduces prediction accuracies.