Machine-learning-based pIC50 prediction identifies novel acetylcholinesterase inhibitors among FDA-approved drugs

Abstract

Acetylcholinesterase (AChE) remains one of the most validated therapeutic targets for the symptomatic treatment of Alzheimer’s disease. Despite decades of research, the repertoire of approved AChE inhibitors is limited, and the identification of novel chemotypes with inhibitory potential is an active area of investigation. We present an automated, reproducible machine-learning pipeline for the prediction of AChE inhibitory activity among approved drugs. Bioactivity records (9731 IC50 measurements) were retrieved from the ChEMBL database for human AChE and preprocessed into a dataset of 6050 compounds. Each molecule was encoded with a dual descriptor set comprising approximately 1600 Mordred 2D physicochemical descriptors and 2048-bit Morgan fingerprints, yielding 3571 features after variance filtering. Two XGBoost models—a regressor for continuous pIC50 prediction and a classifier for binary activity assignment—were independently optimized through Bayesian search over a scaffold-stratified GroupKFold cross-validation scheme to prevent data leakage between related compounds. On a held-out test set of 1076 molecules, the regressor achieved a mean absolute error of 0.661 log units and R2 = 0.642, while the classifier attained an area under the receiver operating characteristic curve of 0.905, an area under the precision-recall curve of 0.895, and an F1 score of 0.796. Model interpretability was assessed via SHapley Additive exPlanations analysis, which highlighted contributions of physicochemical descriptors and topological substructures. An in silico screen of 2062 approved drugs from the DrugBank database identified 256 compounds (12.4%) as predicted active against AChE, including 19 drugs with documented AChE activity in ChEMBL and 235 novel repurposing candidates with no prior AChE record. The pipeline is publicly available as an open-source tool, readily adaptable to other pharmacological targets, facilitating drug-repurposing efforts. Sensitivity analysis confirmed the robustness of the binary activity threshold across alternative cut-offs, and an ablation study demonstrated that the combined Mordred–Morgan feature set yields better cross-validation performance than either descriptor family alone.