Single-step retrosynthesis prediction via multitask graph representation learning.

Area of Science:

• Computational Chemistry
• Organic Synthesis
• Machine Learning

Background:

• Accurate retrosynthesis prediction is crucial for designing novel molecules.
• Existing computational methods for single-step retrosynthesis have limitations in template dependency, interpretability, and entity association utilization.

Purpose of the Study:

• To develop an advanced multitask graph representation learning model for single-step retrosynthesis prediction.
• To simultaneously address reaction center deduction and leaving group identification.

Main Methods:

• Leveraging intra- and inter-associations between synthons and leaving groups (LGs).
• Developing a multitask graph representation learning model named Retro-MTGR.
• Validating model robustness, scalability, and component contributions.

Main Results:

• Retro-MTGR demonstrates superior performance compared to 16 state-of-the-art methods.
• The model can determine potential reaction centers and appropriate leaving groups.
• Inferred retrosynthesis routes are promising for single-step synthesis.

Conclusions:

• Retro-MTGR offers a significant advancement in computational retrosynthesis.
• The model's predictions align with fundamental chemical synthesis principles, particularly regarding electrical properties.
• The developed approach enhances the efficiency and accuracy of predicting chemical synthesis routes.