Prediction of cholinergic compounds by machine-learning

Abstract

The cholinergic nervous system plays a central role in biology and medicine and is comprised of the nicotinic (nAChR) and muscarinic receptors (mAChR) and acetylcholinesterase (AChE). Using ~20,000 compounds compiled from publicly available data, we developed random-forest machine-learning models that distinguish inhibitors of the cholinergic system from non-inhibitors by analyzing 2-D structural scaffolds and fingerprints. We also developed parallel models that incorporate binding energetics data from protein-ligand docking. The scaffold/fingerprint-based models were highly sensitive (98.4–99.3% AChRs; 100% AChE) and accurate (96.1–96.8% AChRs; 100% AChE). By further incorporating the binding energies from protein docking, the models identified compounds that interact with either the nAChR or mAChR with 100% sensitivity while retaining high accuracy (≥91%). Uniquely, these models can predict whether a compound will or will not inhibit the cholinergic system with 91.1–100% balanced accuracy, respectively. In evaluating robustness, ≤12% of available data were needed for highly sensitive models, suggesting that the models will not change for the foreseeable future. Furthermore, we show that training sets for machine learning models requires sufficient diversity of active and inactive (control) compounds for the resulting models to accurately detect a wide array of active substances. Finally, as an example application, when used to query a large in vitro neuro-signaling database, these models flagged all previously identified cholinergic compounds as well as many additional substances with previously uncharacterized mechanisms. These approaches can be readily adopted to model other neuronal and biological protein targets.