Integrating Tox21 Assay Data and Chemical Fingerprints for Neurotoxicity Prediction

Abstract

Exposure to environmental and pharmaceutical chemicals presents a significant risk for neurotoxicity. To reduce reliance on in vivo testing, the U.S. Toxicology in the 21st Century (Tox21) consortium has generated quantitative high-throughput screening data for approximately 10,000 compounds (the Tox21 10K library). This study evaluates the predictive performance of Tox21 assay data and chemical structure information acting simultaneously as input descriptors to determine an independent target variable of potential compound neurotoxicity. We constructed predictive models using five machine learning algorithms: Random Forest, Naïve Bayes, eXtreme Gradient Boosting, Support Vector Machine, and Neural Networks. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC). Two models were built on individual types of chemical fingerprints (ECFP4 and ToxPrint), which achieved AUC-ROC values of 0.72–0.77 and 0.67–0.71, respectively, while those using Tox21 assay data ranged from 0.66–0.73. Notably, models that integrated all three data types outperformed single-source models, achieving AUC-ROC values of 0.78–0.84, indicating that multimodal integration of biological and chemical features reliably enhances predictive accuracy. These models were then applied to virtually screen the Tox21 10K library and to prioritize compounds with the highest predicted neurotoxicity for experimental validation. Key contributors to potential neurotoxicity prediction included assay data on glucocorticoid receptor activation and structural features such as pyridine rings and tertiary aliphatic amines. Overall, this study demonstrates the effectiveness of integrating high-throughput assay data with chemical structure analysis to enable reliable in silico prediction of neurotoxicity, offering a scalable and animal-free approach for assisting chemical safety assessment.