Prediction of Chemical-induced Autonomic Neurotoxicity with Machine Learning Approaches

Autonomic neurotoxicity associated with chemical exposure represents a significant clinical and safety concern. Traditional assessment relies on in vivo methods that are costly and time-consuming, and few computational tools specifically address this endpoint. This study aimed to develop reliable computational models for the prediction of chemical-induced autonomic neurotoxicity, while elucidating the molecular determinants governing this adverse effect.

The autonomic neurotoxicity data were retrieved from the ADReCS and SIDER databases. Multiple machine learning and deep learning algorithms were utilized to develop Quantitative Structure-Activity Relationship (QSAR) models via the ensemble modeling method. The chemical structures were characterized using diverse molecular descriptor packages. Furthermore, molecular properties and structural alerts responsible for autonomic neurotoxicity were analyzed.

The consensus model was built based on four best-performing individual models (Py- Descriptor_RFR, QNPR_RFR, RDKIT_ASNN, and CNF), which achieved good results on both five-fold cross-validation (ACC value of 0.80, AUC value of 0.90, and MCC value of 0.60) and independent test validation (ACC value of 0.84, AUC value of 0.92, and MCC value of 0.68). Additionally, the analysis identified seven key molecular properties (MW, MPSA, AlogP, nHBA, nHBD, nRB, and nAR) and 19 structural alerts that are significantly associated with autonomic neurotoxicity.

The superior performance of the consensus model compared to individual algorithms highlights its ability to enhance the overall stability of the results. The analysis of molecular properties demonstrated that autonomic neurotoxicants tend to exhibit higher lipophilicity, smaller molecular size, reduced hydrogen-bonding capacity, lower molecular flexibility, and increased aromaticity. Furthermore, the 19 structural alerts offer interpretable insights into the specific chemical motifs associated with toxicity. These alerts serve as practical tools for medicinal chemists to facilitate the early identification of toxicophores.

This study comprehensively elucidates the autonomic neurotoxicity and the associated structural hallmarks, offering a validated computational framework to guide highthroughput virtual screening and the safe design of therapeutic agents.