DeepMech: a machine learning framework for chemical reaction mechanism prediction.

Area of Science:

• Computational Chemistry
• Chemical Informatics
• Artificial Intelligence in Chemistry

Background:

• Predicting complete chemical reaction mechanisms (CRMs) is a major scientific challenge.
• Current methods rely on costly experiments or quantum computations, with deep learning (DL) often overlooking crucial intermediates.
• Lack of large, annotated datasets hinders progress in mechanistic prediction.

Purpose of the Study:

• To introduce DeepMech, an interpretable, graph-based DL framework for predicting CRMs step-by-step.
• To address limitations in DL for CRMs by prioritizing reactivity and incorporating mechanistic operations.
• To develop a model trained on a large-scale dataset of annotated reaction mechanisms.

Main Methods:

• Developed DeepMech, a graph-based DL framework using attention mechanisms at atom and bond levels.
• Incorporated a template of mechanistic operations (TMOp) for generating intermediates in elementary steps.
• Constructed ReactMech, a dataset of ~30K full CRMs with ~100K atom-mapped elementary steps for training.

Main Results:

• DeepMech achieved state-of-the-art accuracy: 98.98% for elementary steps and 95.94% for complete CRMs.
• The model demonstrated high fidelity on out-of-distribution data, including unseen catalysts and ligands.
• DeepMech successfully reconstructed complex CRMs relevant to prebiotic chemistry, identifying reactive sites accurately.

Conclusions:

• DeepMech provides a powerful, interpretable tool for data-driven prediction of CRMs.
• The framework has the potential to significantly accelerate mechanistic understanding and reaction design across scientific domains.
• This work advances the field by enabling accurate, step-by-step CRM prediction, overcoming previous DL limitations.