Interpretable Machine Learning-Based Hepatotoxicity Prediction Using Molecular Fingerprints and SHAP-Based Structural Analysis for Animal and Environmental Toxicity Assessment

Abstract

ABSTRACT: Hepatotoxicity is a critical endpoint in toxicological evaluation because exposure to pharmaceuticals, industrial chemicals, agrochemicals, and environmental contaminants can adversely affect liver function in humans and animals. The development of reliable computational approaches for early toxicity prediction can facilitate chemical safety assessment while supporting strategies that reduce reliance on animal testing. In this study, an interpretable quantitative structure–activity relationship (QSAR)-based machine learning framework was developed to predict hepatotoxicity using molecular fingerprints and structural toxicity signatures. Publicly available HepG2 cytotoxicity data from the PubChem BioAssay database and oxidative stress response data from the Tox21 SR-ARE assay were used to construct predictive models based on Morgan fingerprint representations of chemical structures. Three supervised machine learning algorithms—Random Forest, Extreme Gradient Boosting (XGBoost), and Deep Neural Networks—were evaluated using standard classification metrics. Among the tested models, XGBoost achieved the highest predictive performance on the HepG2 dataset, with an area under the receiver operating characteristic curve (AUC-ROC) of 0.9555, demonstrating excellent discrimination between hepatotoxic and non-hepatotoxic compounds. Cross-assay validation using the Tox21 SR-ARE dataset resulted in comparatively lower predictive performance, reflecting differences in biological endpoints between toxicity assays. To improve model interpretability, SHapley Additive exPlanations (SHAP) analysis was employed to identify molecular substructures associated with hepatotoxicity. The analysis identified recurring structural toxicity signatures, including nitrogen-containing heterocyclic systems, electron-withdrawing aromatic substituents, nitrogen-rich aromatic fragments, and sulfur-containing heterocycles such as thiazole and oxazole derivatives. A prototype HepatoRisk Index was further developed by integrating prediction probabilities with structural toxicity signatures to support prioritization of potentially hepatotoxic compounds. The proposed framework provides a transparent, scalable, and computationally efficient approach for preliminary hepatotoxicity prediction and chemical hazard evaluation. By improving the interpretability of computational toxicity models, this study supports safer chemical screening, environmental risk assessment, and the development of alternative predictive methods that complement animal-based toxicological evaluation.