Predicting Active and Inactive Targets of Small Molecules from ChEMBL Data Using the Polypharmacology Browser PPB4

Drug-like molecules often interact with multiple biological targets, which must be identified during drug development. Various models spanning classical machine learning, deep learning and artificial intelligence, predict active targets of a query molecule based on its structural similarity to target annotated molecules from drug-target interaction databases such as ChEMBL. Herein we report on the Polypharmacology Browser PPB4, a deep neural network trained on ChEMBL35. In contrast to its predecessor PPB3 and related tools which only return lists of probably active targets, PPB4 also lists probably inactive targets, as well as targets that are probably both active and inactive if the query molecule is similar to both active and inactive compounds for that target. Two prospective case studies illustrate the use of PPB4, which returns every predicted target together with the list of nearest neighbor molecules marked as active on that target in the training data. PPB4 is freely available at https://ppb4.gdb.tools.