Deep learning-enhanced QSAR modeling for predicting developmental neurotoxicity based on molecular initiating events from adverse outcome pathways

Abstract

Developmental neurotoxicity (DNT) is linked to chemical exposure that disrupts the nervous system in humans or animals. Traditional methods for assessing chemical toxicity are valuable but often time-consuming, costly, and involve significant animal use, making it impractical to meet growing demands. To address this, we developed a deep learning-enhanced QSAR modeling framework aimed at predicting binding affinities towards molecular initiating events (MIEs) and key events (KEs) within the Adverse Outcome Pathway (AOP) relevant to exposure to pesticide-contaminated cannabis. Our model was trained on data from 24,476 compounds, sourced from the ChEMBL database, and tested against 4 MIE and 6 KE tasks. The DNNs showed superior performance, with an average correlation coefficient of 0.82±0.05 and a root mean square error of 0.72±0.08 for the test set. To enhance interpretability, we used SHAP values to explain the model’s predictions clearly. Furthermore, ECFP4 feature contributions were mapped onto known neurotoxic compounds to high- light regions likely responsible for MIEs visually. Our results confirm that developed models accurately predict DNT and effectively identify the correct MIEs and KEs for several neurotoxicants.