Generation of explainable QSAR models for predicting chemical-induced urinary tract toxicity

Area of Science:

• Toxicology
• Computational Chemistry
• cheminformatics

Background:

• Chemical-induced urinary tract toxicity is less understood than nephrotoxicity.
• Existing models lack mechanistic interpretability for urinary tract toxicity.

Purpose of the Study:

• Develop an explainable Quantitative Structure-Activity Relationship (QSAR) framework.
• Predict chemical-induced urinary tract toxicity using mechanistically relevant molecular descriptors.
• Provide a transparent in silico tool for early chemical safety assessment.

Main Methods:

• Utilized a curated dataset of 209 diverse compounds annotated with EPA LD50 thresholds.
• Reduced 1444 molecular descriptors to 11 key variables capturing chemical properties.
• Evaluated 12 machine learning algorithms, with Random Forest showing the best performance (84.13% accuracy, 0.86 AUC).
• Employed SHAP analyses for mechanistic explanations of toxicity drivers.

Main Results:

• The Random Forest model achieved high accuracy and AUC for predicting urinary tract toxicity.
• SHAP analyses identified epithelial barrier disruption, oxidative stress, and altered membrane permeability as key toxicity mechanisms.
• External validation confirmed robust generalization and applicability domain.

Conclusions:

• The developed QSAR framework offers a transparent and mechanistically interpretable tool for early safety evaluation.
• This approach supports risk assessment, safer-by-design chemistry, and regulatory decision-making.
• The tool promotes the 3Rs (Replacement, Reduction, Refinement) by minimizing animal testing.