Pathway-Aware Template-Based Retrosynthesis

Area of Science:

• Computational chemistry
• Machine learning in chemical synthesis

Background:

• Machine learning accelerates the discovery of synthetic routes for high-value compounds like drugs.
• Current single-step retrosynthesis models often lack pathway context, causing search inefficiencies.

Purpose of the Study:

• To introduce the PATRO model, which incorporates pathway-level information into template-based retrosynthesis.
• To evaluate PATRO's performance against baseline models in both single-step and multistep retrosynthesis.

Main Methods:

• Augmented a template-based single-step retrosynthesis model using a Long Short-Term Memory (LSTM) network.
• Processed pathway-level information to inform reaction predictions.
• Integrated the pathway-aware model into two multistep retrosynthesis algorithms.

Main Results:

• PATRO improved single-step retrosynthesis top-1 accuracy by 2.3% compared to the baseline.
• Consistent performance gains were observed in multistep retrosynthesis metrics (success rate, patent route recovery, top-k accuracy).
• PATRO reduced planning iterations by 10% for discovering literature routes from patents.

Conclusions:

• Incorporating pathway-level information significantly enhances retrosynthesis model performance.
• The PATRO model offers a more efficient approach to chemical synthesis planning.
• This strategy holds potential for improving various other retrosynthesis models.