Predictive Artificial Intelligence Approaches for Accurate Drug Classification in Pharmaceutical Industry

Abstract

Drug classification plays a critical role in medicine as it aids in selecting the best medicines for an individual’s needs based on their individual characteristics and history. Computational methods are increasingly used in drug discovery to build structure-activity models for large chemical databases. This study introduces a scalable ML approach for drug classification based on scaffolds using SMILES from the ChEMBL database. The approach uses RDKit for physicochemical descriptor extraction, Bemis-Murcko scaffolds for target construction and employs feature selection, encoding, RobustScaler normalization and SMOTE for balancing classes. Models include Random Forest, XGBoost, and a Stacking Ensemble, with accuracy, precision, recall, F1-score, and ROC-AUC as evaluation metrics. The experimental findings show that the Stacking Ensemble outperforms Random Forest (88.06%) and XGBoost (85.91%), achieving an accuracy of 89.15% and a ROC-AUC of 98.29%, suggesting better generalization. The results demonstrate that ensemble and tree-based learning methods outperform traditional models and LSTM in scaffold classification. The novel approach provides a rapid, scalable, and precise framework that increases the efficiency of virtual screening and offers a reliable approach to AI-based decision-making in drug discovery.