Chemistry-informed molecular graph as reaction descriptor for machine-learned retrosynthesis planning

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry

Background:

• Data-driven retrosynthesis planning often lacks chemical intuition.
• Integrating chemical knowledge can improve the accuracy of predicting synthetic pathways.

Purpose of the Study:

• To develop a novel approach for automatic retrosynthesis planning by incorporating chemical wisdom.
• To design a chemistry-informed molecular graph (CIMG) representation for chemical reactions.

Main Methods:

• A chemistry-informed molecular graph (CIMG) was designed, incorporating NMR chemical shifts, bond dissociation energies, and solvent/catalyst information.
• Graph neural networks (GNNs) were employed to predict reaction templates, catalysts, and solvents based on CIMG representations.
• Pretrained reaction vectors derived from CIMGs were used to accelerate Monte Carlo tree search (MCTS) for multistep retrosynthesis.

Main Results:

• The CIMG-based approach successfully generated plausible reaction vectors containing chemical wisdom.
• The method demonstrated efficiency in predicting full synthetic routes with recommended catalysts and solvents.
• Autocategorization of molecules and reactions was achieved by exploiting reaction propensity information.

Conclusions:

• Infusing chemical wisdom into data-driven retrosynthesis planning significantly improves prediction accuracy and efficiency.
• The CIMG framework provides a robust method for representing chemical reactions and guiding retrosynthesis.
• This approach accelerates complex multistep synthesis planning and aids in reaction condition selection.