Virtual screening using local neuro-fuzzy rules

As an application of a neuro-fuzzy approach we present results of drug target molecules classification. Inactive molecules are separated from active ones for different ligand data sets. Our technique can be seen as a retrospective virtual screening method. As a basis the molecule data is encoded in descriptor vectors. We compare two descriptors, one encoding two-dimensional topological features and one encoding three-dimensional distances of atom types. ROC area for classification and enrichment factors of active molecules in local rules are compared. Although one could assume that 3D descriptors contain the more performance features than the 2D ones, we show that the used 2D descriptor has superior performance for the considered datasets.