NeuroTDPi: Interpretable Deep Learning Models with Multimodal Fusion for Identifying Neurotoxic Compounds

Area of Science:

• Computational toxicology
• cheminformatics
• machine learning in drug discovery

Background:

• Chemical neurotoxicity is a significant concern in drug development and environmental safety.
• Early prediction of neurotoxicity can reduce experimental costs and accelerate safety assessments.
• Current methods may lack the accuracy and efficiency needed for large-scale screening.

Purpose of the Study:

• To develop and validate NeuroTDPi, a deep neural network model for predicting chemical neurotoxicity.
• To integrate multimodal data for enhanced prediction accuracy across multiple neurotoxicity endpoints.
• To improve the interpretability of neurotoxicity predictions and identify structural alerts.

Main Methods:

• Developed NeuroTDPi, a multilayer fully connected deep neural network.
• Employed a multimodal fusion strategy integrating molecular characterization and endpoint-specific features.
• Utilized SHapley Additive Explanations (SHAP) for model interpretability and identification of key chemical properties.
• Evaluated model performance using area under the receiver operating characteristic curve (AUC).

Main Results:

• NeuroTDPi achieved high predictive performance with AUC values of 0.97 (blood-brain barrier permeability), 0.84 (neuronal toxicity), and 0.82 (mammalian neurotoxicity).
• The SHAP analysis provided insights into the physical and chemical properties influencing neurotoxicity predictions.
• Identified specific structural alerts associated with neurotoxic compounds, offering mechanistic understanding.

Conclusions:

• NeuroTDPi offers a robust and interpretable platform for early-stage neurotoxicity evaluation.
• The model facilitates risk assessment by providing actionable structural insights for chemical safety.
• Freely available resources at https://www.sapredictor.cn/ support broader application in toxicology and drug development.