ChronicDPipredictor: an interpretable deep learning framework for chemical chronic and subchronic toxicity...

Area of Science:

• Computational toxicology
• Cheminformatics
• Machine learning for drug safety

Background:

• Evaluating long-term (chronic) and medium-term (subchronic) chemical toxicity is crucial but challenging due to complex mechanisms and diverse structures.
• Developing accurate predictive computer models for these toxicity endpoints remains a significant hurdle in chemical safety assessment.

Purpose of the Study:

• To develop an interpretable machine learning framework, ChronicDPipredictor, for assessing chemical chronic and subchronic toxicity.
• To enhance the interpretability of toxicity predictions using SHAP analysis.
• To provide a publicly accessible web-server for toxicity prediction and identify structural alerts associated with toxicity.

Main Methods:

• Developed ChronicDPipredictor using machine learning with MACCS, PubChem, and KRFP fingerprints.
• Evaluated model performance in three-class and binary classification for chronic and subchronic toxicity.
• Applied SHAP analysis for model interpretability and extracted structural alerts from KRFP fingerprints.

Main Results:

• Models based on MACCS fingerprints achieved the highest performance, with accuracies up to 0.82 (chronic) and 0.80 (subchronic) in three-class classification.
• Binary classification accuracy reached 0.93 for chronic and 0.83 for subchronic toxicity.
• Identified 18 structural alerts for chronic toxicity and 7 for subchronic toxicity, with several linked to known toxicological mechanisms.

Conclusions:

• ChronicDPipredictor offers an accurate and interpretable approach for assessing chemical chronic and subchronic toxicity.
• The framework and its identified structural alerts aid in the risk assessment of compound repeated-dose toxicity.
• The developed web-server provides a practical tool for researchers and regulators.