A chemical class-based approach to predictive model generation

We make a quantitative comparison of two distinct approaches to predictive model generation in the context of diverse screening data. In the default approach, a single recursive partitioning model is constructed using all of the training data at one time. In the "class-based" approach, the same data are first partitioned into homogeneous, scaffold-based classes, and models are constructed within each class independently. Both approaches are tested on the identical set of hold-out data, using a formal protocol that includes consensus scoring to handle the multiple class-based models. The entire process is performed using three different descriptor sets and is repeated using five separate random trials, such that the trial-averaged prediction rates for the two approaches can be quantitatively compared. We find that although the predictive performances of the class-based and default approaches are similar, the former has at least two distinct advantages. The first is greater interpretability, in that chemists can more easily extract useful structure-activity information from the models. The second is greater reliability, allowing models to be applied with increased confidence to unseen data in virtual-screening applications.