Machine Learning-Aided Drug Repurposing for Screening COX-2 Inhibitors from Traditional Chinese Medicines.

Background/Objectives: Machine learning has emerged as a transformative force in drug discovery, revolutionizing traditional research paradigms and profoundly improving the efficiency, cost-effectiveness, and speed of the drug development cycle for novel drugs. Colorectal cancer is one of the most prevalent malignant tumors and imposes a heavy burden on global public health due to its high morbidity, mortality, and poor prognosis. Cyclooxygenase-2 (COX-2) is a key therapeutic target of colorectal cancer and has been extensively applied in the development of novel anti-colorectal cancer drugs. Methods: In this study, we systematically compared the performance of Random Forest Classifier (RFC), deep learning (DL), and graph neural network (GNN) models, including GAT (Graph Attention Network), GCN (Graph Convolutional Network), and MPNN (Message Passing Neural Network), with diverse features in the classification task of COX-2 inhibitors, based on a custom COX-2 inhibitors dataset and a Traditional Chinese Medicine (TCM)-derived compound library. The optimal model was subsequently used to screen for potential COX-2 inhibitors. Additionally, the key substructures governing COX-2 inhibitory activity were also identified and analyzed. Finally, the prioritized candidate compounds underwent experimental validation. Results: Both RFC and DL models outperformed GNN models. Through further comparative analysis of models' predictive performance, the RFC model was ultimately verified as the optimal model for activity screening of TCM-derived compounds. The molecular interactions and binding affinities between predicted candidate compounds and COX-2 were further investigated. Finally, the selected lead compound, dehydrocostus lactone, was experimentally confirmed to possess potent COX-2 inhibitory activity. Conclusions: This study highlights that the RFC model is highly effective in screening bioactive components from TCM under small-dataset conditions, providing a solid foundation for subsequent related research in this field.