J Chem Inf Model. 2026 Feb 20. doi: 10.1021/acs.jcim.6c00299. Online ahead of print.
ABSTRACT
Drug-like molecules often interact with multiple biological targets. Assessing this polypharmacology is essential for drug development. Here, we trained deep neural networks to associate bioactive molecules up to 80 non-hydrogen atoms reported in ChEMBL 34, represented as binary substructure fingerprints, with lists of targets on which the molecules are ≥50% active at ≤10 μM. We included 2,496,555 interactions between 1,187,089 molecules and 7546 targets having at least five reported active molecules, including single proteins, protein complexes, protein families, cell lines, organisms, and further target types. This represents a much larger data set than in previously reported models, which were mostly limited to protein targets. Our models achieve good performances in terms of recall and precision per molecule and per target, as illustrated by overall statistics and by a case study in comparison with other online prediction tools. PPB3 predictions can be performed online at https://ppb3.gdb.tools/.
PMID:41721463 | DOI:10.1021/acs.jcim.6c00299
