Employing Machine Learning Models to Predict Potential α-Glucosidase Inhibitory Plant Secondary Metabolites Targeting Type-2 Diabetes and Their In Vitro Validation.

Abstract

The need for new antidiabetic drugs is evident, considering the ongoing global burden of type-2 diabetes mellitus despite notable progress in drug discovery from laboratory research to clinical application. This study aimed to build machine learning (ML) models to predict potential α-glucosidase inhibitors based on the data set comprising over 537 reported plant secondary metabolite (PSM) α-glucosidase inhibitors. We assessed 35 ML models by using seven different fingerprints. The Random forest with the RDKit fingerprint was the best-performing model, with an accuracy (ACC) of 83.74% and an area under the ROC curve (AUC) of 0.803. The resulting robust ML model encompasses all reported α-glucosidase inhibitory PSMs. The model was employed to predict potential α-glucosidase inhibitors from an in-house 5810 PSM database. The model identified 965 PSMs with a prediction activity ≥0.90 for α-glucosidase inhibition. Twenty-four predicted PSMs were subjected to in vitro assay, and 13 were found to inhibit α-glucosidase with IC₅₀ ranging from 0.63 to 7 mg/mL. Among them, seven compounds recorded IC₅₀ values less than the standard drug acarbose and were investigated further to have optimal drug-likeness and medicinal chemistry characteristics. The ML model and in vitro experiments have identified nervonic acid as a promising α-glucosidase inhibitor. This compound should be further investigated for its potential integration into the diabetes treatment system.