Random Forest Model for the Prediction of Herbal-Induced Liver Injury: Application to Molecules from the West African Pharmacopoeia

Abstract

Herbal medicines play a crucial role in primary healthcare across West Africa, yet their potential for liver toxicity remains poorly documented. Predicting herbal-induced liver injury is therefore essential to ensure the safe use of traditional remedies. A Random Forest model was developed to predict hepatotoxicity using a combined feature set of nine physicochemical descriptors and Morgan fingerprints R2 (1024 bits). The training dataset included reference compounds from the FDA DILIst dataset, while external validation was performed on the Greene dataset. Model performance was assessed using nested cross-validation and evaluated through multiple metrics including AUC-ROC, AUC-PR, F1-macro, and MCC. The combined descriptors-molecular fingerprints model achieved an external AUC-ROC of 0.83, AUC-PR of 0.85, and MCC of 0.55, demonstrating strong generalization capacity. Application of the model to 191 phytochemicals from the West African pharmacopoeia indicated that 80.1% were potentially hepatotoxic, with fusidic acid and nicamin showing the highest probabilities (>80%). These results confirm the reliability of the RF approach for hepatotoxicity prediction and highlight the need for systematic toxicological evaluation of traditional medicines. Artificial intelligence thus offers an efficient framework for integrating safety assessment into pharmacopoeia modernization.