Enhancing diversity in language based models for single-step retrosynthesis.

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry

Background:

• Natural Language Processing (NLP) models are increasingly used in scientific research.
• Retrosynthesis, a key chemical challenge, has seen success with NLP but often lacks diverse predictions.
• Current models propose precursors from limited reaction families, restricting chemical space exploration.

Purpose of the Study:

• To develop a retrosynthesis Transformer model that improves the diversity of predicted molecular precursors.
• To enable broader exploration of chemical space by overcoming limitations in current retrosynthesis approaches.

Main Methods:

• A Transformer model was adapted for retrosynthesis, treating it as a translation problem.
• Classification tokens were prepended to the language representation of target molecules.
• These prompt tokens were used at inference to guide the model toward diverse disconnection strategies.

Main Results:

• The proposed model consistently improved the diversity of predicted retrosynthesis pathways.
• The use of classification tokens effectively steered the model to generate varied disconnection strategies.
• Enhanced diversity allows recursive synthesis tools to avoid dead ends.

Conclusions:

• The novel retrosynthesis Transformer model significantly enhances prediction diversity.
• This approach facilitates the suggestion of synthesis pathways for more complex molecules.
• The method offers a promising direction for advancing AI-driven chemical synthesis.