Interpretable machine learning for predicting key hazardous properties of chemicals.

Area of Science:

• Computational Chemistry
• Chemical Informatics
• Predictive Modeling

Background:

• Accurate prediction of chemical hazards (toxicity, flammability, reactivity, reactivity with water) is crucial for safety.
• Experimental methods are time-consuming, costly, and limited in capturing complex structure-property relationships.
• Existing models often lack interpretability, hindering understanding of molecular drivers.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting four key hazardous chemical properties.
• To enhance model interpretability for understanding structure-property relationships.
• To apply the optimal model for predicting hazards in a list of chemicals.

Main Methods:

• Development of 8 machine learning models using molecular descriptors.
• Training on self-curated datasets with advanced feature selection and interpretability techniques.
• Utilizing XGBoost and Random Forest (RF) models, with SHAP and ICE for analysis.

Main Results:

• XGBoost excelled in predicting toxicity (ROC-AUC 0.768) and reactivity (0.917).
• RF demonstrated superior performance for flammability (0.952) and reactivity with water (0.852).
• Key molecular descriptors (MIC4, ATSC2i, ATS4i, ETA_dEpsilon_C) identified for specific hazards in Ketone/Aldehyde compounds.

Conclusions:

• Machine learning models provide efficient and scalable predictions of chemical hazards.
• These models can reduce reliance on costly experimental testing.
• The study enhances safety protocols for hazardous chemical management through predictive insights.