RetroCaptioner: beyond attention in end-to-end retrosynthesis transformer via contrastively captioned learnable graph...

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry

Background:

• Retrosynthesis is crucial for identifying precursors for novel compounds.
• Transformer models are increasingly used for automated retrosynthesis.
• Existing methods often fail to capture reaction transformation details, limiting accuracy.

Purpose of the Study:

• To introduce RetroCaptioner, an advanced end-to-end Transformer-based framework for chemical retrosynthesis.
• To improve the accuracy and applicability of automated retrosynthesis predictions.
• To effectively capture reaction transformation information and chemically plausible constraints.

Main Methods:

• Developed RetroCaptioner, featuring a Contrastive Reaction Center Captioner for training dual-view attention models.
• Utilized contrastive learning with molecular graph representations for a single-step learning process.
• Integrated single-encoder, dual-encoder, and encoder-decoder paradigms, enhancing atomic correspondence between SMILES and molecular graphs.

Main Results:

• Achieved top-1 accuracy of 67.2% and top-10 accuracy of 93.4% on the USPTO-50k dataset.
• Obtained an exceptional SMILES validity score of 99.4%.
• Demonstrated reliability in generating synthetic routes for the drug protokylol.

Conclusions:

• RetroCaptioner significantly advances automated chemical retrosynthesis.
• The framework's ability to capture reaction details and constraints leads to high accuracy and validity.
• This method shows promise for accelerating drug discovery and development.