It is crucial to identify the molecular targets of a compound during the course of the new
drug discovery and drug development. Due to the complexity of biological systems, finding drug
targets by biological experiments is very tedious and expensive. In the paper, we used chemicalchemical
interactions in the STITCH database to construct a network of drug-drug association. Based
on the network, a learning method keeping local and global consistency was presented to infer drug targets. We achieved
an accuracy of 57.75% in the first order prediction using leave-one-out cross validation, which was higher than the
accuracy of 53.77% achieved by the local neighbor model. We manually validated 27 absent drug targets in the crossvalidation
using drug-target interactions from other databases. Applying the presented method to large-scale prediction of
unknown targets, we manually confirmed 14 pairs of drug-target interactions among the newly predicted drug targets.
These results suggested that the presented method was a promising tool for large-scale identification of drug targets.
Keywords: Drug target, chemical-chemical interaction, semi-supervised learning, leave-one-out cross validation, guilt by
association, in silico prediction.
Rights & PermissionsPrintExport