Pharmacophore features for machine learning in pharmaceutical virtual screening

Methods of three-dimensional molecular alignment generally treat all pharmacophore features equally when superimposing. However, some pharmacophore features can be more important in a specific system. In this work, we derived the overlap volume of pharmacophore features from a molecular alignment approach as new features of molecules to build machine learning models. Features can be assigned weights to indicate their importance. With validation on DUD-E collection, models based on pharmacophore features represented by the overlap volume yielded significant performances with median AUC of approximately 0.98 and recall rate of almost 0.8.