Evaluation of Nonparametric Models for Genomic Prediction of Early-Stage Single Crosses in Maize

Prediction of single-cross performance in a hybrid breeding program is extremely important because it is not feasible to evaluate all parental combinations. Recent simulation and field studies have shown great promise of genomic prediction of single-cross performance. These previous studies, however, have primarily focused on parametric genomic prediction models. This study tested three nonparametric models-reproducing kernel Hilbert spaces, support vector regression, and neural networks-for prediction of early-stage single crosses. Two separate datasets, consisting of 481 and 312 single crosses, were used to evaluate models. Single crosses were made by randomly crossing inbred progenies between heterotic groups. Genomic prediction models were trained to directly predict single-cross performance, or to predict general combining ability (GCA) of inbred parents and specific combining abilities (SCA) of single crosses between them. Using cross-validation, genomic predictions were compared with predictions using phenotypes of single crosses with a common parent (common-parent single crosses), as well as phenotypic estimates of GCA. Of these three options for predicting single-cross performance, genomic prediction resulted in the highest correlation between observed and predicted values. Predictive abilities of parametric and nonparametric genomic prediction models were nearly identical. All genomic prediction models displayed good ability to predict GCA effects, but none could predict SCA effects. Our results suggest that nonparametric models do not provide an advantage over parametric models for prediction of early-stage, single-cross performance using modestly sized training populations like those used here.