Development of machine learning models to predict acetylcholinesterase inhibitory activity and UDP-glucuronosyltransferase affinity and design of new acetylcholinesterase inhibitors

Although acetylcholinesterase (AChE) inhibitors help slow the progression of Alzheimer's disease (AD), proper drug metabolism is required to minimize side effects. Drug elimination occurs mainly through phase I and phase II metabolism, mediated by cytochrome P450 (CYP) and uridine diphosphate (UDP)-glucuronosyltransferase (UGT), respectively. Considering that CYP activity declines with aging, identifying compounds with high UGT affinity is important for safe drug clearance. In this study, molecular descriptors and machine learning models were used to predict AChE inhibitory activity (pIC 50 ) and UGT affinity (pK m ). The best models achieved predictive performance of r 2 = 0.54 for AChE inhibition and r 2 = 0.77 for UGT affinity. Using these models, novel compounds were designed via a generative adversarial network (GAN), and several candidates exhibited higher predicted pIC 50 and pK m values than donepezil (pIC 50 = 6.09, pK m = 4.21). Molecular docking confirmed greater binding stability for the generated structures, suggesting their potential as effective and metabolically favorable AChE inhibitors. This approach demonstrates that integrating predictive modeling and generative design can accelerate the discovery of safer AD treatments. • Machine learning models were built to predict AChE inhibition and UGT affinity. • RDKit-PLS and FP&RDKit-GPR models showed highest prediction performance. • GAN-based design identified novel compounds exceeding known AChE and UGT values. • Designed molecules demonstrated stronger binding than donepezil via molecular docking. • The approach offers a data-driven pipeline for designing safer AChE inhibitors.