Modelling Chemical Reasoning to Predict and Invent Reactions

Area of Science:

• Organic Chemistry
• Artificial Intelligence
• Cheminformatics

Background:

• Organic chemists require reasoning beyond established knowledge for synthetic problem-solving.
• Predicting novel chemical reactions is a significant challenge in synthetic chemistry.

Purpose of the Study:

• To develop a computational model that mimics chemical reasoning for reaction prediction.
• To formalize reaction prediction as a knowledge graph link prediction task.

Main Methods:

• Construction of a large-scale knowledge graph with 14.4 million molecules and 8.2 million reactions.
• Development of a data-driven model to predict missing links (reactions) within the graph.
• Evaluation of the model against a rule-based expert system using 180,000 random binary reactions.

Main Results:

• The proposed model demonstrated superior performance in reaction prediction compared to a rule-based expert system.
• The data-driven model successfully generalized beyond known reaction types, identifying novel transformations.
• The model can generate reaction hypotheses rapidly (sub-second per prediction), facilitating high-throughput reaction discovery.

Conclusions:

• The AI model effectively replicates chemical reasoning for predicting organic reactions.
• The knowledge graph approach enables the discovery of novel chemical transformations, including transition metal-catalyzed reactions.
• This high-throughput prediction capability can accelerate the discovery of new synthetic routes.