PPB4 deep network predicts both active and inactive targets for small molecules from ChEMBL35
- PPB4 is a deep neural network trained on ChEMBL35 that infers likely targets of a query molecule from its similarity to target-annotated compounds.
- Unlike the earlier PPB3 and comparable tools that report only probable actives, it also flags likely inactive targets, and targets where the query resembles both active and inactive compounds.
- Each predicted target comes with its nearest active neighbors from the training data. Validation is shown through two prospective case studies; no comparative benchmark is described. Free web server.
💡 Points Based on the abstract
Flagging likely inactives is the useful bit: ChEMBL inactives are sparse and assay-dependent, so I'd want to know what threshold defines 'inactive' and how calibrated those calls are. With no head-to-head against PPB3, SEA or similar, and only two case studies, I can't tell if it's better. I'd run it on a few of our compounds with known off-targets before trusting it.