DeepHeptox: An Interpretable Deep Learning Model for Multi-Endpoint Hepatotoxicity Prediction of Chemical Compounds

Area of Science:

• Toxicology and Computational Chemistry
• Drug Discovery and Development
• Bioinformatics and Cheminformatics

Background:

• Hepatotoxicity is a significant concern in chemical exposure, impacting drug development and safety.
• The liver's role in metabolism necessitates accurate hepatotoxicity prediction.
• Computational methods offer an efficient and ethical alternative to experimental toxicity testing.

Purpose of the Study:

• To develop a deep learning model for predicting multi-endpoint hepatotoxicity.
• To identify structural alerts and visualize molecular importance for hepatotoxicity.
• To provide a user-friendly web server for interactive toxicity prediction.

Main Methods:

• Development of DeepHeptox, a Graph Attention Network (GAT) based deep learning model.
• Multi-endpoint prediction covering hepatitis, jaundice, elevated liver enzymes, hepatocellular injury, hepatic fibrosis, hepatomegaly, and cholestasis.
• Utilized Klekota-Roth fingerprint (KRFP) analysis for structural alert identification and GNNExplainer for visualization.

Main Results:

• DeepHeptox achieved high predictive performance with Area Under the ROC Curve (AUC) > 0.87 and Accuracy (ACC) > 0.80.
• The model demonstrated effectiveness in both overall and endpoint-specific hepatotoxicity predictions.
• Structural alerts and molecular substructure importance were successfully identified and visualized.

Conclusions:

• DeepHeptox provides a practical and accurate computational tool for early-stage hepatotoxicity screening.
• The model supports drug development and chemical safety assessment by predicting liver injury.
• The integrated web server facilitates interactive analysis and visualization of toxicity predictions.