Evaluation of prediction methods with increasing amounts of randomly selected training data. Ten random sequences of experiments were used to select data used for training regression models. After each experiment was chosen, a ROC curve was constructed by gradually raising the threshold on the predicted assay score at which an experiment was considered to be positive. The mean and standard error of the area under the ROC curve for prediction of positive experiments after each experiment is plotted for each regression method. The prediction methods shown are: within target random prediction in red, regression using protein features only (PFO, cyan), regression using compound features only (CFO, blue) and CCT (green).