Empirical testing of chemicals for drug efficacy costs many billions of dollars every year. The ability to predict the action of molecules in silico would greatly increase the speed and decrease the cost of prioritizing drug leads. Here, we asked whether drug function, defined as MeSH ‘Therapeutic Use’ classes, can be predicted from only chemical structure. We evaluated two chemical structure-derived drug classification methods, chemical images with convolutional neural networks and molecular fingerprints with random forests, both of which outperformed previous predictions that used drug-induced transcriptomic changes as chemical representations. This suggests that a chemical’s structure contains at least as much information about its therapeutic use as the transcriptional cellular response to that chemical. Further, because training data based on chemical structure is not limited to a small set of molecules for which transcriptomic measurements are available, our strategy can leverage more training data to significantly improve predictive accuracy to 83-88%. Finally, we explore use of these models for prediction of side effects and drug repurposing opportunities, and demonstrate the effectiveness of this modeling strategy for multi-label classification.

Virtual (computational) high-throughput screening provides a strategy for prioritizing compounds for experimental screens, but the choice of virtual screening algorithm depends on the dataset and evaluation strategy. We consider a wide range of ligand-based machine learning and docking-based approaches for virtual screening on two protein-protein interactions, PriA-SSB and RMI-FANCM, and present a strategy for choosing which algorithm is best for prospective compound prioritization. Our workflow identifies a random forest as the best algorithm for these targets over more sophisticated neural network-based models. The top 250 predictions from our selected random forest recover 37 of the 54 active compounds from a library of 22,434 new molecules assayed on PriA-SSB. We show that virtual screening methods that perform well on public datasets and synthetic benchmarks, like multi-task neural networks, may not always translate to prospective screening performance on a specific assay of interest.