A benchmark dataset and interpretable deep learning framework for drug-induced developmental neurotoxicity prediction

Developmental neurotoxicity (DNT) represents a critical yet underevaluated toxicity endpoint within current chemical safety assessment frameworks, particularly in the context of drug development. In the present study, we constructed a standardized benchmark dataset comprising 2724 structurally curated compounds, including 1362 DNT-positive chemicals derived from the Developmental Neurotox Reference List (DNTREF) of the U.S. Environmental Protection Agency (EPA) and 1362 property-matched presumed negative compounds (non-neuroactive drug-like compounds) selected from the ChEMBL database. Using this benchmark dataset, we systematically evaluated seven representative quantitative structure activity relationship (QSAR) modeling paradigms, covering traditional machine learning, deep neural networks, molecular language models, and graph neural networks. Across all models, the area under the receiver operating characteristic curve (AUC) values ranged from 0.85 to 0.90 on an independent test set. Among these, a deep neural network based on MACCS fingerprints achieved the highest sensitivity in identifying DNT-positive compounds, supporting its utility in early-stage drug safety screening where minimizing false negatives is critical. Furthermore, SHAP-based interpretability analysis revealed key structural motifs associated with DNT, highlighting the roles of hydrophobic aromatic frameworks, electrophilic reactivity, and polar functional groups in modulating predicted neurodevelopmental risk. Collectively, this study provides a reproducible benchmark dataset and an interpretable QSAR framework that supports early DNT risk prioritization and offers actionable structural insights for medicinal chemistry optimization in drug discovery.