Virtual screening encompasses a wide range of computational approaches aimed at the
high-throughput, cost-efficient exploration of chemical libraries or databases to discover new bioactive
compounds or novel medical indications of known drugs. Here, we have performed a systematic comparison of the
performance of a large number of 2D and 3D ligand-based approaches (2D and 3D similarity, QSAR models,
pharmacophoric hypothesis) in a simulated virtual campaign on a chemical library containing 50 known anticonvulsant
drugs and 950 decoys with no previous reports of anticonvulsant effect. To perform such comparison, we resorted to
Receiver Operating Characteristic curves. We also tested the relative performance of consensus methodologies. Our
results indicate that the selective combination of the individual approaches (through voting and ranking combination
schemes) significantly outperforms the individual algorithms and/or models. Among the best-performing individual
approaches, 2D similarity search based on circular fingerprints and 3D similarity approaches should be highlighted.
Combining the results from different query molecules also led to enhanced enrichment.
Keywords: 2d fingerprints, comparison, consensus scoring, data fusion, ligand-based virtual screening, molecular similarity.
Rights & PermissionsPrintExport