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

Neurotoxicity from drugs and environmental pollutants poses serious risks to brain function, yet existing computational models mainly target general neurotoxicity and lack specialized tools for brain-specific assessment. This study aimed to develop and validate a high-performance, brain-focused neurotoxicity prediction framework to improve drug safety evaluation and toxicity screening. We systematically analyzed molecular features, clustering patterns, and target predictions of brain-toxic compounds. Multiple feature representations were compared, including traditional molecular fingerprints, knowledge-guided pre-trained graph Transformer (KPGT) embeddings, and transformer-based MolFormer embeddings, combined with machine learning classifiers. Model performance was evaluated using multiple metrics, and SHAP analysis was conducted to identify influential molecular substructures. Toxic molecules showed physicochemical properties favoring central nervous system (CNS) penetration, including lower molecular weight, lower LogP, fewer hydrogen bond donors/acceptors, fewer rotatable bonds, and lower polar surface area (PSA). The KPGT-MLP model achieved the best balanced performance, with an accuracy (ACC) of 0.8928 and an ROC-AUC of 0.9459, clearly outperforming traditional fingerprint-based models, MolFormer-based models, and general prediction tools such as DI-NeuroT and ADMETlab 3.0. Overall, this study establishes a robust framework for brain-specific neurotoxicity prediction, with the KPGT-MLP model demonstrating strong accuracy and robustness. The proposed approach provides an effective strategy for early neurotoxicity screening and risk assessment, offering valuable insights for safer drug design and advancing computational toxicology and drug discovery.