Generative Modeling to Predict Multiple Suitable Conditions for Chemical Reactions

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Synthetic Chemistry

Background:

• Accurate prediction of chemical reaction conditions is crucial for successful synthesis planning.
• Machine learning methods have shown promise in recommending reaction elements but offer limited predictions.
• Existing models provide single predictions, lacking complete specification of reaction conditions.

Purpose of the Study:

• To develop a generative modeling approach for predicting multiple feasible reaction conditions for chemical reactions.
• To address the limitations of existing methods by providing complete reaction condition specifications.
• To enhance the performance in retrieving ground-truth reaction conditions.

Main Methods:

• Formulated reaction condition prediction as sampling from a generative distribution.
• Employed a variational autoencoder augmented with a graph neural network.
• Trained the model on a reaction dataset, enabling multiple predictions via repeated sampling.

Main Results:

• The proposed generative model successfully predicts multiple, fully specified reaction conditions.
• Experimental validation on cross-coupling reaction datasets demonstrated superior performance compared to existing methods.
• Significantly improved retrieval of ground-truth reaction conditions.

Conclusions:

• The generative modeling approach offers a significant advancement in predicting chemical reaction conditions.
• This method provides a more comprehensive and accurate solution for synthesis planning.
• The approach holds potential for broader applications in computational chemistry and drug discovery.