Predicting retrosynthetic pathways using transformer-based models and a hyper-graph exploration strategy.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry

Background:

• Automated retrosynthesis planning is crucial for efficient chemical synthesis.
• Existing methods often require significant human intervention and lack comprehensive evaluation.

Purpose of the Study:

• To develop an end-to-end framework for automatic retrosynthesis route planning.
• To establish novel metrics for evaluating single-step retrosynthetic models.

Main Methods:

• Extension of the Molecular Transformer model with a hyper-graph exploration strategy.
• Development of four metrics: coverage, class diversity, round-trip accuracy, and Jensen-Shannon divergence.
• On-the-fly hyper-graph construction with Bayesian-like probability-based node filtering and expansion.

Main Results:

• The single-step retrosynthetic model achieves state-of-the-art performance in predicting reactants, reagents, solvents, and catalysts.
• The end-to-end framework demonstrates excellent performance on literature and academic examples.
• Identified minor weaknesses related to training data limitations.

Conclusions:

• The integrated AI framework enables fully automated retrosynthesis route planning.
• The novel metrics allow for focused optimization of retrosynthetic models.
• The approach shows significant potential for accelerating chemical synthesis and discovery.