Assessing Prediction Models for Different Traits in a Rice Population Derived from a Recurrent Selection Program

Genomic selection (GS) is a promising approach to improve rice (Oryza sativa L.) populations by using genome-wide markers for selection prior to phenotyping to estimate breeding values. In this study, our objectives were to compare certain prediction models with different structures of genetic relationship and statistical approaches for relevant traits in rice and to discuss some implications for integrating GS into a recurrent selection program of irrigated rice. We assessed nine models in terms of predictive potential, using empirical data from S-1:3 progenies phenotyped for eight traits with different heritabilities and genotyped with 6174 high-quality single nucleotide polymorphism markers. For all traits, marker-based models outperformed prediction based on pedigree records alone. A similar level of accuracy was observed for many models, although the level of prediction stability and prediction bias varied widely. Random forest was slightly superior for less complex traits, although with high prediction bias, whereas the semiparametric RKHS method (reproducing kernel Hilbert spaces) was superior for many traits, showing high stability and low bias. Bayesian variable selection method Bayes C pi showed acceptable accuracy and stability for several traits and thus could be useful for genomic prediction aiming at persisting accuracy for a long-term recurrent selection.