Prediction of Chemical-induced Autonomic Neurotoxicity with Machine Learning Approaches

Area of Science:

• Computational toxicology
• cheminformatics
• drug discovery

Background:

• Chemical exposure poses risks of autonomic neurotoxicity, a significant safety concern.
• Current in vivo assessment methods are costly and time-consuming.
• Limited computational tools exist for predicting chemical-induced autonomic neurotoxicity.

Purpose of the Study:

• Develop reliable computational models for predicting chemical-induced autonomic neurotoxicity.
• Identify molecular determinants associated with this adverse effect.
• Provide tools for early identification of toxicophores in drug design.

Main Methods:

• Utilized machine learning and deep learning algorithms for Quantitative Structure-Activity Relationship (QSAR) model development.
• Employed ensemble modeling and diverse molecular descriptor packages for model building.
• Analyzed molecular properties and structural alerts linked to autonomic neurotoxicity using ADReCS and SIDER databases.

Main Results:

• A consensus model achieved high accuracy (ACC=0.84, AUC=0.92) in predicting autonomic neurotoxicity.
• Identified seven key molecular properties (e.g., lipophilicity, molecular size) associated with toxicity.
• Discovered 19 structural alerts indicative of autonomic neurotoxic potential.

Conclusions:

• The consensus model demonstrates superior performance and stability for predicting chemical-induced autonomic neurotoxicity.
• Autonomic neurotoxicants exhibit specific molecular characteristics, including higher lipophilicity and aromaticity.
• Structural alerts provide actionable insights for medicinal chemists to design safer therapeutic agents.