HazChemNet: A Deep Learning Model for Hazardous Chemical Prediction

Area of Science:

• Computational Chemistry
• Toxicology
• Artificial Intelligence

Background:

• Accurate identification of hazardous chemicals is crucial for environmental and health protection.
• Current methods for chemical hazard assessment often suffer from inefficiency and lack of precision.

Purpose of the Study:

• To develop and validate HazChemNet, a novel deep learning model for predicting chemical hazardousness.
• To leverage AI for enhanced chemical safety and risk management in industrial practices.

Main Methods:

• Developed HazChemNet, a deep learning model combining attention-based autoencoders and mixture-of-experts architectures.
• Utilized a dataset of 2428 hazardous compounds from China's 2015 list, extracting features from molecular fingerprints and physicochemical descriptors.
• Performed external validation on 52 unseen chemicals and experimental validation using *C. elegans* assays.

Main Results:

• HazChemNet achieved 92.3% accuracy in classifying hazardous chemicals and 84.6% for non-hazardous ones during external validation.
• Experimental validation using *C. elegans* assays corroborated the model's predictions for critical compounds.
• Ablation studies identified hydrogen bonding features and molecular fingerprints as key predictors.

Conclusions:

• HazChemNet offers a transformative AI-driven tool for accurate and efficient chemical hazard identification.
• The model supports sustainable industrial practices and proactive risk management in chemical safety.
• This work integrates artificial intelligence with chemical safety protocols for a safer global landscape.