3DReact: Geometric Deep Learning for Chemical Reactions.

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Informatics

Background:

• Geometric deep learning models enhance molecular property prediction by integrating molecular symmetries.
• Accurate prediction of chemical reaction properties is crucial for drug discovery and materials science.

Purpose of the Study:

• To introduce 3DReact, a novel geometric deep learning model for predicting chemical reaction properties.
• To evaluate 3DReact's performance on predicting activation barriers using 3D molecular structures.

Main Methods:

• Developed 3DReact, a geometric deep learning model utilizing 3D reactant and product structures.
• Employed invariant and equivariant neural network architectures.
• Tested on GDB7-22-TS, Cyclo-23-TS, and Proparg-21-TS datasets across different atom-mapping regimes.

Main Results:

• The invariant version of 3DReact demonstrated sufficient performance for existing reaction datasets.
• Achieved competitive accuracy in predicting activation barriers.
• Showcased robust performance across diverse datasets, atom-mapping strategies, and prediction tasks (interpolation/extrapolation).

Conclusions:

• 3DReact provides a flexible and effective framework for predicting chemical reaction properties.
• The model successfully leverages 3D structural information and atom-mapping data.
• Demonstrates systematic and strong performance, outperforming existing models in reaction property prediction.