DiffER : categorical diffusion ensembles for single-step chemical retrosynthesis

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry

Background:

• Automatic chemical retrosynthesis leverages natural language processing (NLP) models, particularly transformers.
• Current transformer models excel at translating chemical structures (SMILES) but are limited by their autoregressive nature.

Purpose of the Study:

• Introduce DiffER, a novel template-free method for single-step retrosynthesis prediction.
• Explore categorical diffusion for simultaneous prediction of the entire output SMILES sequence.

Main Methods:

• Developed an ensemble of diffusion models for retrosynthesis.
• Implemented a template-free approach using categorical diffusion.

Main Results:

• Achieved state-of-the-art top-1 accuracy in retrosynthesis prediction.
• Demonstrated competitive top-3, top-5, and top-10 accuracy compared to other template-free methods.
• Showcased DiffER's ability to learn diverse laboratory synthetic techniques.

Conclusions:

• DiffER establishes a strong baseline for a new class of template-free retrosynthesis models.
• The categorical diffusion approach offers advantages over autoregressive methods.
• DiffER shows promise for practical application in chemical synthesis planning.