Epistasis and Improved Prediction of Performance in Corn

To be useful, adding epistatis to a prediction model must increase predictive power. The objectives of this study were to determine: 1) using partial least squares techniques, whether the ability to predict performance can be increased by including epistasis in a prediction model; 2) whether relaxing the probability level for inclusion of a marker or interaction in a model from 0.001 to 0.01 to 0.05 would increase predictive power; 3) whether the proportion of variability accounted for could be raised to a level useful in breeding; 4) whether molecular marker models based on one year’s phenotypic data could be used to predict performance in a second year; and 5) whether models based on per se data would be useful in predicting testcross performance. Data for protein, oil, and starch, were obtained from 500 S2 lines and their testcrosses from the crosses of Illinois High Oil (IHO) x Illinois Low Oil (ILO) and of Illinois High Protein (IHP) x Illinois Low Protein (ILP) corn (Zea mays L.) strains. Increasing the probability level for detection of significant markers and epistatic effects from 0.001 to 0.01 to 0.05 significantly increased predictive power. Adding epistasis to a model significantly increased predictive power when models were based on data over all year and locations, when models based on one year were used to predict a second year, or when per se data were used to predict testcross performance. With epistasis in the model and P=0.05, correlations of predicted and observed means were high enough to suggest they might be useful in breeding. Marker models based on one year’s data produced correlations nearly as high as phenotypic correlations between years. While epistasis significantly improved performance of models used to predict testcross performance from per se performance, the proportion of variability accounted for was somewhat lower than when predicting performance in different years.