Discriminating of HMG-CoA reductase inhibitors and decoys using self-organizing maps.

Self-Organizing Map (SOM) models were built to distinguish inhibitors of HMG-CoA reductase from its non-binding decoys. The molecules were represented by five global molecular descriptors and seven 2D property autocorrelation descriptors. Based on these molecular descriptors, 35 HMG-CoA reductase ligands and 1480 decoys were projected into a self-organizing network. In the map, the ligands and the decoys were well separated, where no neuron was occupied by a ligand and a decoy at the same time. Afterward, the discriminating power of the selected molecular descriptors was further validated by extending the datasets to 135 inhibitors. Finally, the SOM approach was subsequently used to identify active compounds in a virtual screening experiment by an external test set which included 32 HMG-CoA reductase inhibitors and 1103 decoys. In this study, 84.4% of the inhibitors (true positives) are retrieved with 15% contamination by non-hits (false positives). The SOM models obtained in this article exhibited powerful ability in virtual screening to find novel inhibitors for HMG-CoA reductase.