Segregating confident predictions of chemicals' properties for virtual screening of drugs

Abstract

In this paper we present a methodology for evaluating the confidence in the prediction of a physicochemical or biological property. Identifying unreliable compounds' predictions is crucial for the modern drug discovery process.This task is accomplished by the combination of the method of prediction with a self-organizing map. In this way, the method is able to segregate unconfident predictions as well as confident predictions. We applied the method to four different data sets, and we obtained significant differences in the average predictions of our segregation. This approach constitutes a novel way for evaluating confidence, since it not only looks for extrapolation situations but also it identifies interpolation problems. © 2009 Springer Berlin Heidelberg.