Chemically Informed Deep Learning for Interpretable Radical Reaction Prediction.

Area of Science:

• Chemistry
• Computational Chemistry
• Chemical Reaction Prediction

Background:

• Organic radical reactions are fundamental in synthetic, biological, and atmospheric chemistry.
• Predicting the outcomes of these reactions is essential for understanding and controlling chemical processes.

Purpose of the Study:

• To develop a predictive framework for mechanistic-level organic radical reactions.
• To provide interpretable predictions based on molecular orbital interactions.

Main Methods:

• A chemistry-aware model was developed, focusing on molecular orbital interactions.
• The model was trained and evaluated using the RMechDB database of radical reaction steps.
• A pathway search was implemented by chaining model predictions.

Main Results:

• The model achieved 96% accuracy in predicting correct orbital interactions and products on the RMechDB test set.
• The pathway search successfully identified intermediates and byproducts in atmospheric and polymerization chemistry problems.
• The RMechRP tool is available online for public use.

Conclusions:

• The developed framework offers accurate and interpretable predictions for organic radical reactions.
• The pathway search capability enhances the understanding of complex radical reaction mechanisms.
• This approach has significant implications for advancing chemical synthesis, atmospheric science, and polymerization studies.