Machine learning-driven prioritization and experimental validation of traditional Chinese medicine-derived...

The stimulator of interferon genes (STING) is a key signalling adaptor in the cGAS-STING pathway of the innate immune system and plays a significant role in autoimmune diseases, viral infections, and cancer, thus representing a promising target for small-molecule inhibitor therapies. This study presents an integrated multidimensional computer-aided drug design (CADD) approach that utilises machine learning (ML), molecular docking, molecular dynamics (MD) simulations, and ADMET prediction to efficiently prioritize new STING expression suppressor candidates from natural products. We developed a precise ML-based STING classification model with 90.2% accuracy and a robust STING inhibitor activity regression model demonstrating strong predictive capabilities, as evidenced by an R² of 0.826, MAE of 0.357, and RMSE of 0.452. Virtual screening across multiple traditional Chinese medicine (TCM) compound libraries (Tao Shu L6810, TCMIO, TCMBank, and HERB) yielded 1,596 compounds with predicted pIC₅₀ ≥ 7.00. After a rigorous multistep screening, seven compounds were selected for ADMET evaluation and experimental validation. Notably, two natural compounds, Cassiaside and Plantaginin, showed inhibitory activity on STING protein expression in THP-1-derived macrophages, and MD simulations, along with CETSA experiments, further validated their stable binding to the STING protein. Collectively, this study provides a robust and accurate ML-driven strategy for STING-Inhibitory Candidates discovery and prioritized two promising TCM-derived lead compounds that offer valuable structural scaffolds for the rational design of STING-targeted therapeutics against immune and inflammatory diseases.