Establishment of an Integrated Model for Predicting Compound Mutagenicity with a Feature Importance Analysis

Area of Science:

• Computational chemistry and toxicology
• Artificial intelligence in chemical safety assessment

Background:

• Assessing chemical mutagenicity is vital for public health and environmental safety.
• Traditional methods like the Ames test are slow and costly for large-scale screening.
• Deep learning offers a faster, more cost-effective approach to mutagenicity prediction.

Purpose of the Study:

• To develop and evaluate an integrated deep learning framework for predicting compound mutagenicity.
• To identify optimal molecular features and model combinations for accurate mutagenicity assessment.
• To provide insights into structural features associated with mutagenic potential.

Main Methods:

• Developed 78 integrated deep learning models by combining 13 types of molecular descriptors and fingerprints.
• Trained models on 5279 compounds and evaluated on 587 compounds.
• Conducted activity cliff and applicability domain analyses to assess model reliability and identify misprediction sources.

Main Results:

• The MACCS-Mordred deep learning model achieved the highest performance with 0.885 balanced accuracy and 0.922 precision.
• Applicability domain analysis confirmed the model's robustness for the tested compounds.
• Feature importance analysis highlighted nitrogen-containing and ring substructures as key indicators of mutagenicity.

Conclusions:

• AI-enabled deep learning models provide a powerful tool for rapid and cost-effective mutagenicity screening.
• The developed framework enhances early-stage chemical risk assessment and aids in prioritizing hazardous compounds.
• Findings support the use of AI for environmental monitoring and regulatory decision-making.