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

Nephrotoxicity remains a critical safety concern in drug development and clinical practice. Despite their significance, existing computational models for nephrotoxicity prediction face challenges related to limited precision and reliability. To address these challenges, this study constructed the largest publicly available database to date, comprising 1831 high-quality nephrotoxicity-related compounds. Using this dataset, we developed classification models employing both traditional machine learning algorithms and graph-based deep learning methods. Our results demonstrate that the Directed Message Passing Neural Network model, combined with molecular graphs and ChemoPy2D descriptors, outperformed other models, achieving a mean Kappa value of 70.3%. To improve the reliability of the model, we implemented an uncertainty quantification method to define the model's applicability domain and quantify the confidence of prediction. On the representative model, this approach significantly enhanced predictive performance within the applicability domain, yielding a Kappa metric of 90.4%. Notably, our model also achieved the highest performance on an external test set compared to existing models. Finally, we performed multiscale feature analysis to provide actionable insights for safer drug design. This analysis integrated dataset -centric methods to identify structural alerts and beneficial transformations, alongside model-centric techniques to reveal the key global and local features driving nephrotoxicity. The established prediction models, combined with uncertainty quantification and structural feature insights derived in this study, offer valuable tools for the development of safer and more effective drugs.