Node-Aligned Graph-to-Graph: Elevating Template-free Deep Learning Approaches in Single-Step Retrosynthesis.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Artificial Intelligence

Background:

• Deep learning (DL) models enhance computer-aided synthesis design for retrosynthesis.
• Template-free DL models for retrosynthesis often neglect 2D molecular information and atom alignment, limiting performance.
• Existing methods struggle with accuracy compared to template-based approaches.

Purpose of the Study:

• Introduce node-aligned graph-to-graph (NAG2G), a novel transformer-based template-free DL model.
• Address limitations of current template-free DL models in retrosynthesis prediction.
• Improve accuracy and robustness in predicting chemical synthesis pathways.

Main Methods:

• Developed NAG2G, a transformer-based template-free DL model.
• Integrated 2D molecular graphs and 3D conformations for comprehensive molecular representation.
• Incorporated product-reactant atom mapping via node alignment for autoregressive node generation.

Main Results:

• NAG2G demonstrated remarkable predictive accuracy on USPTO-50k and USPTO-FULL datasets.
• The model successfully predicted synthesis pathways for drug candidate molecules.
• Achieved superior performance compared to existing template-free DL methods.

Conclusions:

• NAG2G offers a robust and accurate solution for template-free retrosynthesis prediction.
• The model's ability to handle complex synthesis processes shows potential for revolutionizing synthetic route design.
• NAG2G advances the field of computer-aided synthesis design with its innovative approach.