Valid, Plausible, and Diverse Retrosynthesis Using Tied Two-Way Transformers with Latent Variables

Area of Science:

• Computational Chemistry
• Organic Synthesis
• Artificial Intelligence in Chemistry

Background:

• Retrosynthesis is crucial for discovering new material synthesis pathways in organic chemistry.
• Deep learning, particularly transformer models, shows promise for automating retrosynthesis.
• Existing pure transformer models yield results lacking chemical validity and diversity.

Purpose of the Study:

• To develop an improved deep learning model for chemical retrosynthesis.
• To address limitations of pure transformer models in accuracy, validity, and diversity.
• To enhance the identification of plausible reactant candidates for synthesis.

Main Methods:

• Development of a tied two-way transformer model incorporating latent variable modeling.
• Implementation of cycle consistency checks and parameter sharing for improved performance.
• Utilization of multinomial latent variables to enhance diversity and validity.

Main Results:

• The proposed model significantly improves retrosynthesis accuracy compared to pure transformers.
• Demonstrated reduction in grammatical errors and enhanced chemical plausibility of results.
• Increased diversity in suggested reactant candidates, providing broader synthesis options.

Conclusions:

• The tied two-way transformer with latent modeling effectively addresses limitations in deep learning-based retrosynthesis.
• The model generates chemically valid, plausible, and diverse synthesis pathway suggestions.
• This approach represents a significant advancement for computational organic chemistry and materials discovery.