Machine Learning Assisted QSAR Modelling of Human-MAO-B Inhibitors

Monoamine oxidase-B (MAO-B) is one of the valid therapeutic targets for the effective treatment of neurodegenerative disorders, which include both Parkinson's disease and Alzheimer's disease. In this study, a machine-learning-based quantitative structure–activity relationship (QSAR) model was developed to predict the inhibitory activity of MAO-B inhibitors. A dataset of 1614 compounds was retrieved from the BindingDB database. After being subjected to comprehensive data preprocessing that included the removal of duplicates and invalid SMILES, resulting in a final dataset of 1612 compounds. Using RDKit, around 219 2D molecular descriptors were calculated that was further reduced to 131 during preprocessing by eliminating near-constant and highly correlated features. The dataset was further divided into training (80%) and test (20%) sets for external validation. Multiple baseline regression algorithms, including Linear Regression, Ridge Regression, LASSO, K-Nearest Neighbours, Random Forest, Extra Trees, Gradient Boosting Decision Trees, and Support Vector Regression, were evaluated. Advanced ensemble learning approaches, including Stacking, Voting Regressor, LightGBM, XGBoost, and CatBoost, were further employed to improve predictive performance. The robustness of the model was assessed using external validation and 5-fold cross-validation, with the optimal model achieving a mean R² of approximately 0.71, demonstrating good predictive capability. The developed QSAR model provides a reliable prediction of MAO-B inhibitory activity and supports the virtual screening and rational optimization of novel MAO-B inhibitors for neurodegenerative disease drug discovery