Uncertainty-Aware Deep Learning and Structural Feature Analysis for Reliable Nephrotoxicity Prediction.

Area of Science:

• Computational toxicology
• Drug discovery and development
• Medicinal chemistry

Background:

• Nephrotoxicity is a major safety concern in drug development, with current predictive models lacking precision.
• Existing computational methods for nephrotoxicity prediction face limitations in reliability and accuracy.

Purpose of the Study:

• To construct the largest publicly available database of nephrotoxicity compounds.
• To develop and validate advanced machine learning and deep learning models for nephrotoxicity prediction.
• To enhance model reliability through uncertainty quantification and provide insights for safer drug design.

Main Methods:

• Compiled a dataset of 1831 high-quality nephrotoxicity-related compounds.
• Developed classification models using traditional machine learning and graph-based deep learning (Directed Message Passing Neural Network).
• Integrated uncertainty quantification and multiscale feature analysis for model validation and insight generation.

Main Results:

• The Directed Message Passing Neural Network model achieved a mean Kappa of 70.3%.
• Uncertainty quantification enhanced predictive performance within the applicability domain to a Kappa of 90.4%.
• The developed model demonstrated superior performance on an external test set compared to existing methods.

Conclusions:

• The developed models, incorporating uncertainty quantification, offer a significant advancement in predicting drug-induced kidney toxicity.
• Multiscale feature analysis provides actionable insights for designing safer drugs with reduced nephrotoxic potential.
• This work establishes valuable tools for improving drug safety and efficacy in pharmaceutical research.