Transfer Learning for Heterocycle Retrosynthesis.

Area of Science:

• Medicinal Chemistry
• Organic Synthesis
• Computational Chemistry

Background:

• Heterocycles are vital scaffolds in drug design, influencing binding and pharmacokinetics.
• Existing datasets lack synthetic routes, hindering access to novel heterocyclic compounds.
• Current retrosynthesis models perform poorly on heterocycle formation due to limited data.

Purpose of the Study:

• To improve retrosynthesis prediction performance for heterocycle formation reactions.
• To address the challenge of low data availability in heterocyclic synthesis.
• To enhance the accessibility of novel and uncommon heterocyclic scaffolds.

Main Methods:

• Comparison of four transfer learning methods for retrosynthesis prediction.
• Application of transfer learning to overcome data scarcity in heterocycle synthesis.
• Development of a mixed fine-tuned model for ring-breaking disconnections.

Main Results:

• The mixed fine-tuned model achieved 36.5% top-1 accuracy.
• 62.1% of the model's predictions were chemically valid and ring-breaking.
• Demonstrated model applicability by recreating synthetic routes for two drug-like targets.

Conclusions:

• Transfer learning significantly improves retrosynthesis prediction for heterocycles.
• The developed model enhances the synthetic accessibility of novel heterocyclic compounds.
• A method for continuous model improvement with new reaction data is introduced.