Improving Few- and Zero-Shot Reaction Template Prediction Using Modern Hopfield Networks.

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Drug discovery and materials science

Background:

• Computer-assisted synthesis planning (CASP) is crucial for discovering new drugs and materials.
• Current CASP methods often rely on single-step models of chemical reactivity.
• Improving the efficiency and accuracy of retrosynthesis prediction is an ongoing challenge.

Purpose of the Study:

• To develop a novel template-based single-step retrosynthesis model.
• To leverage Modern Hopfield Networks for enhanced prediction of template relevance.
• To accelerate the process of finding synthesis routes for molecules.

Main Methods:

• Implementation of a template-based retrosynthesis model utilizing Modern Hopfield Networks.
• Learning joint encodings of molecules and reaction templates.
• Predicting the relevance of chemical reaction templates for given molecules.

Main Results:

• The proposed model significantly improves template relevance prediction accuracy, especially for data-sparse templates.
• Achieved state-of-the-art performance in top-k exact match accuracy (k>=3) on the USPTO-50k retrosynthesis benchmark.
• Demonstrated inference speeds orders of magnitude faster than baseline methods.

Conclusions:

• The Modern Hopfield Network-based approach offers a significant advancement in computer-assisted synthesis planning.
• This method enhances the generalization capability of retrosynthesis models across diverse reactions.
• The developed tool accelerates the identification of potential synthesis routes, aiding drug and material discovery.