Identification and Classification of GPCR Ligands Using Self-Organizing Neural Networks
Paul M. Selzer, Peter F. Ertl · QSAR & Combinatorial Science · 2005
A combination of a 3D descriptor and neural networks was applied to model the relationship between molecular structures and their activity as G-Protein-Coupled Receptor, GPCR ligands. The 3D descriptor, namely the Radial Distribution Function, RDF, is based on the interatomic distances and thus expresses the molecule's pharmacophoric features. The first goal of the study was to analyze whether the RDF code provides sufficient GPCR relevant structural information to separate GPCR ligands from a set of randomly selected molecules. Cluster experiments with neural networks show a clear separation of these two classes and even a separation between different classes of GPCR ligands. In a second series of experiments neural networks were used to predict a GPCR-ligand-likeness score. Based on this score 71% of the GPCR ligands (in a data set with only 5.9% active compounds) were correctly identified in a cross-validation experiment.