Neurotoxicity Prediction of Compounds: Integrating Knowledge-Guided Graph Representations with Machine Learning...

Area of Science:

• Computational toxicology
• Neuroscience
• Drug discovery

Background:

• Neurotoxicity from drugs and pollutants threatens brain function.
• Existing computational models lack brain-specific neurotoxicity assessment.
• Need for specialized tools in drug safety evaluation and toxicity screening.

Purpose of the Study:

• Develop and validate a high-performance, brain-focused neurotoxicity prediction framework.
• Improve drug safety evaluation and toxicity screening through specialized assessment.
• Enhance computational toxicology for safer drug design.

Main Methods:

• Systematic analysis of molecular features, clustering, and target predictions for brain-toxic compounds.
• Comparison of feature representations: molecular fingerprints, knowledge-guided pre-trained graph Transformer (KPGT) embeddings, and MolFormer embeddings.
• Machine learning classifiers and SHAP analysis for model evaluation and substructure identification.

Main Results:

• Brain-toxic molecules exhibit physicochemical properties favoring central nervous system (CNS) penetration.
• The KPGT-MLP model achieved superior performance (ACC: 0.8928, ROC-AUC: 0.9459) over other models and general tools.
• Identified influential molecular substructures contributing to brain-specific toxicity.

Conclusions:

• Established a robust framework for brain-specific neurotoxicity prediction.
• The KPGT-MLP model demonstrates high accuracy and robustness for neurotoxicity assessment.
• Provides an effective strategy for early neurotoxicity screening, risk assessment, and safer drug design.