Screening of Respiratory Toxicity of Environmental Compounds Based on Multimodal Feature Fusion Model

Area of Science:

• Computational toxicology
• Environmental health
• Regulatory science

Background:

• The respiratory system is vulnerable to chemical toxicants, but current assessment methods are limited.
• Existing regulatory frameworks rely on expensive, low-throughput animal testing and lack premarket evaluation.
• There is a need for efficient, systematic methods for respiratory toxicity assessment.

Purpose of the Study:

• To develop an innovative multimodal deep learning (DL) framework, GFEnet, for respiratory toxicity prediction.
• To integrate diverse molecular features for comprehensive cross-scale toxicity assessment.
• To establish a high-throughput screening platform for early identification of respiratory toxicants.

Main Methods:

• Developed GFEnet, a multimodal DL framework integrating molecular graph features, structural fingerprints, and electron-level properties.
• Trained and evaluated GFEnet on in vivo mammalian respiratory toxicity, in vitro A549 cell cytotoxicity, and ACE2 gene regulation.
• Validated GFEnet predictions with in vivo mouse models.

Main Results:

• GFEnet achieved high predictive performance with AUC values of 0.986, 0.965, and 0.919.
• Ablation studies confirmed the importance of each feature modality.
• GFEnet identified fluorene-9-bisphenol and Michler's ketone as high-risk candidates, confirmed by mouse model studies.

Conclusions:

• GFEnet is a robust, high-throughput screening platform for early respiratory toxicant identification.
• The framework effectively integrates computational toxicology with environmental health and regulatory science.
• GFEnet advances the assessment of chemical risks to respiratory health.