Developing a multi-task quantitative structure-activity relationship (QSAR) model for predicting the toxicity...

Area of Science:

• Computational toxicology
• Environmental health
• Chemical risk assessment

Background:

• E-cigarette emissions contain numerous compounds with limited toxicity data.
• This data gap hinders hazard identification and chemical prioritization for e-cigarette constituents.

Purpose of the Study:

• To develop and evaluate machine learning (ML)-based quantitative structure-activity relationship (QSAR) models for predicting e-cigarette toxicity.
• To assess the performance of single-task versus multi-task learning strategies for toxicity prediction.
• To identify key chemical features associated with specific toxicity endpoints.

Main Methods:

• Developed interpretable ML-based QSAR models for five toxicity endpoints: eye irritation, skin irritation, skin sensitization, carcinogenicity, and genotoxicity.
• Implemented both single-task and multi-task learning approaches.
• Analyzed feature importance to understand structure-toxicity relationships.

Main Results:

• Multi-task QSAR models achieved moderate to high performance for eye/skin toxicity (AUC ≥ 0.69) and carcinogenicity/genotoxicity (AUC ≥ 0.66).
• Significant performance improvements were observed for eye irritation prediction using multi-task learning.
• Feature analysis indicated nitrogen-containing groups are key for irritation toxicity, while electrostatic properties influence carcinogenicity.

Conclusions:

• Multi-task QSAR modeling effectively predicts toxicity of understudied e-cigarette chemicals.
• This approach can elucidate relationships between different toxicity endpoints.
• Supports model-informed chemical hazard evaluation for e-cigarette products.