A data-driven group retrosynthesis planning model inspired by neurosymbolic programming

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry
• Drug discovery informatics

Background:

• Deep generative models accelerate drug discovery but face challenges in synthesizing proposed molecules.
• Current retrosynthetic planning methods often process molecules independently, missing reusable synthesis patterns.
• AI-generated small molecules present unique challenges due to novel structures and synthesis pathways.

Purpose of the Study:

• To develop an advanced retrosynthetic planning algorithm that leverages reusable synthesis patterns.
• To enhance the efficiency and accuracy of predicting reactions in retrosynthesis search.
• To improve the validation of molecules generated by deep generative models.

Main Methods:

• Developed a neurosymbolic programming-inspired algorithm with wake, abstraction, and dreaming phases.
• Augmented the reaction template library with reusable synthesis patterns discovered from data.
• Implemented an evolutionary process to refine prediction models for reaction templates.
• Applied the algorithm to groups of similar molecules to identify shared synthesis routes.

Main Results:

• The algorithm identified and incorporated reusable synthesis patterns, reducing marginal inference time.
• The neurosymbolic approach demonstrated superior performance compared to existing retrosynthesis methods.
• Significant reductions in inference time were observed when planning retrosynthesis for similar molecules.
• The method effectively discovers underlying chemistry patterns and enhances model prediction.

Conclusions:

• The proposed algorithm offers a more efficient and effective approach to retrosynthetic planning.
• This method has the potential to significantly accelerate the drug discovery pipeline by improving molecule synthesis validation.
• The evolutionary learning process allows the model to adapt and improve over time, discovering novel chemical insights.