Predicting potentially hazardous chemical reactions using an explainable neural network.

Area of Science:

• Chemical Safety
• Artificial Intelligence
• Machine Learning

Background:

• Predicting hazardous chemical reactions is crucial for laboratory safety but is time-consuming and costly via traditional experiments.
• Existing machine learning models for predicting major reaction products lack the required accuracy for identifying hazardous byproducts, posing safety risks due to high false negative rates.

Purpose of the Study:

• To develop an explainable artificial intelligence (AI) model for accurately predicting the formation of hazardous chemical reaction products.
• To achieve a low false negative rate for enhanced laboratory safety, complementing existing prediction models and experimental methods.

Main Methods:

• Developed an explainable AI model using a convolutional neural network (CNN).
• Encoded reactant molecules into substructure-encoded fingerprints as input for the CNN.
• Applied layer-wise relevance propagation (LRP) for input attribution to explain model predictions.

Main Results:

• The proposed model achieved a false negative rate of 0.09, a significant improvement over existing main product prediction models (0.47-0.66).
• Input attribution analysis using LRP provided chemical insights, aligning with known mechanisms and highlighting potential data imbalance issues.
• The model demonstrated the capability to identify hazardous product formation with high accuracy and provide interpretable explanations.

Conclusions:

• The explainable AI model offers a more accurate and reliable method for predicting hazardous chemical reactions, enhancing laboratory safety.
• The model's ability to provide explanations aids in understanding prediction rationales and addressing data limitations.
• This approach serves as a valuable complement to traditional experimental methods and existing machine learning tools for chemical safety.