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

Area of Science:

• * Computational toxicology and cheminformatics
• * Drug safety and developmental neurotoxicity assessment

Background:

• * Developmental neurotoxicity (DNT) is an under-evaluated toxicity endpoint in drug development.
• * Current safety assessment frameworks require improved methods for DNT risk evaluation.

Purpose of the Study:

• * To construct a standardized benchmark dataset for DNT prediction.
• * To evaluate various quantitative structure-activity relationship (QSAR) modeling paradigms for DNT.
• * To identify structural features associated with DNT risk.

Main Methods:

• * Creation of a benchmark dataset with 2724 compounds (1362 DNT-positive, 1362 presumed negative).
• * Systematic evaluation of seven QSAR modeling paradigms, including machine learning and deep learning.
• * Application of SHAP for interpretability analysis of DNT-associated structural motifs.

Main Results:

• * QSAR models achieved an area under the receiver operating characteristic curve (AUC) of 0.85–0.90 on an independent test set.
• * A deep neural network using MACCS fingerprints demonstrated high sensitivity for identifying DNT-positive compounds.
• * Key structural motifs linked to DNT risk were identified, including hydrophobic aromatic frameworks and electrophilic groups.

Conclusions:

• * The study provides a reproducible benchmark dataset and an interpretable QSAR framework for DNT risk prioritization.
• * Deep learning models are effective for early-stage drug safety screening to minimize false negatives.
• * Actionable structural insights are offered for medicinal chemistry optimization in drug discovery.