Single-step retrosynthesis prediction by leveraging commonly preserved substructures

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Retrosynthesis analysis is crucial for organic chemistry and industry.
• Current machine learning models use string representations, limiting human interpretability.
• Expert chemists analyze reactions based on stable molecular substructures.

Purpose of the Study:

• To develop a novel machine learning model for retrosynthesis analysis that decodes at the substructure level.
• To enable chemists to gain more intuitive insights into reaction mechanisms.
• To improve the accuracy and applicability of computational retrosynthesis.

Main Methods:

• Developed a data-driven approach to automatically extract commonly preserved substructures from product molecules.
• Implemented a substructure-level decoding model for predicting reactant molecules.
• Evaluated model performance against existing atom-level decoding methods.

Main Results:

• The substructure-level decoding model demonstrated improved performance compared to previous models.
• Model performance was further enhanced by increasing the accuracy of extracted substructures.
• The model provides interpretable substructural insights for chemists.

Conclusions:

• Substructure-level analysis offers a more interpretable and effective approach to computational retrosynthesis.
• The developed model enhances decision-making for chemists in reaction planning.
• Future work can focus on refining substructure extraction for further performance gains.