AiZynthTrain: Robust, Reproducible, and Extensible Pipelines for Training Synthesis Prediction Models

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Chemical synthesis prediction

Background:

• Retrosynthesis is crucial for planning chemical synthesis.
• Existing methods for training synthesis models often lack reproducibility and extensibility.
• Developing robust and adaptable tools for computational chemistry is essential.

Purpose of the Study:

• Introduce the AiZynthTrain Python package for training chemical synthesis models.
• Provide reproducible pipelines for template-based retrosynthesis and RingBreaker models.
• Demonstrate the package's robustness and potential for future extensions.

Main Methods:

• Developed two training pipelines within the AiZynthTrain package.
• Utilized a publicly available reaction dataset from the U.S. Patent and Trademark Office (USPTO).
• Implemented new heuristics to enhance the RingBreaker model's performance on ring system disconnection.

Main Results:

• Created the first end-to-end reproducible retrosynthesis models.
• The RingBreaker model showed significantly improved performance in disconnecting ring systems.
• Demonstrated pipeline robustness through training on a diverse proprietary dataset.

Conclusions:

• AiZynthTrain provides a robust, reproducible, and extensible framework for training synthesis models.
• The implemented heuristics enhance the capability of retrosynthesis prediction, particularly for complex ring systems.
• The framework is poised for future integration of additional synthesis models.