Predicting Retrosynthetic Reactions Using Self-Corrected Transformer Neural Networks

Area of Science:

• Chemistry
• Artificial Intelligence
• Computational Chemistry

Background:

• Computer-aided retrosynthesis aids chemists in designing synthetic routes.
• Current methods are often cumbersome and yield unsatisfactory results.

Purpose of the Study:

• To develop an accurate and efficient template-free retrosynthesis prediction tool.
• To improve upon existing deep learning and template-based retrosynthesis methods.

Main Methods:

• Developed a template-free self-corrected retrosynthesis predictor (SCROP).
• Utilized transformer neural networks, framing retrosynthesis as a machine translation problem.
• Integrated a neural network-based syntax corrector for enhanced prediction.

Main Results:

• Achieved 59.0% accuracy on a standard benchmark dataset.
• Outperformed other deep learning methods by over 21% and template-based methods by over 6%.
• Demonstrated 1.7 times higher accuracy than state-of-the-art methods for novel compounds.

Conclusions:

• SCROP offers a significant advancement in computer-aided retrosynthesis.
• The template-free, self-corrected approach enhances prediction accuracy and applicability.
• This method shows great promise for accelerating chemical synthesis design.