Hepatotoxicity prediction for traditional Chinese medicine: a two-step in silico framework integrating network and machine learning approaches.

Traditional Chinese medicine (TCM) has gained increasing attention due to several severe cases of herb-induced liver injury over the last few decades. Due to the intricate components and complex internal interactions, the identification of hepatotoxic ingredients in TCM remains a significant challenge. In this study, we proposed a novel two-step in silico framework, which integrates a network-based systems pharmacology approach and a machine learning-based consensus model, to identify potential hepatotoxicity from TCM-derived compounds. Compared to currently available tools, our method showed superior predictive capability in terms of accuracy (0.76) and F1 score (0.72) on a natural product-specific benchmark test set. Based on the two-step in silico framework, we conducted a comprehensive screening of 3,882 TCM compounds, revealing that 133 exhibited potential high-risk hepatotoxicity, including 45 (33.8%) corroborated by clinical and experimental evidence. Cell viability tests and serum biochemistry analyses on HepG2 cells confirmed the significant hepatotoxicity of the predicted compounds, including dehydroevodiamine, monocrotaline, tetrahydrocoptisine, citric acid, and eugenol. We next performed pharmacovigilance assessment of hepatotoxicity for the top 100 commonly used TCM herbs in clinical practice, prioritized the high-risk herbs, and explored the hepatotoxic mechanisms of TCM. Finally, we selected Rheum palmatum L. (Dahuang) as case study to showcase how the proposed two-step in silico framework aids in accurately identifying the compositions and understanding the underlying molecular mechanisms of TCM-induced liver injury. Overall, this study demonstrates a potent computational toxicology framework for TCM-induced hepatotoxicity prediction, aiming to provide guidance for the clinical use of TCM.