Accelerating the inference of string generation-based chemical reaction models for industrial applications

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry

Background:

• Transformer models achieve high accuracy in chemical reaction prediction and retrosynthesis.
• Slow inference speeds hinder the practical application of these models in computer-aided synthesis planning.

Purpose of the Study:

• To accelerate inference speeds for transformer-based SMILES-to-SMILES translation models.
• To enhance the utility of these models in reaction prediction and retrosynthesis for synthesis planning.

Main Methods:

• Applied speculative decoding with a chemically specific drafting strategy to the Molecular Transformer.
• Introduced Speculative Beam Search to accelerate the generation of multiple precursor SMILES in single-step retrosynthesis.

Main Results:

• Achieved over 3X faster inference for reaction product prediction and single-step retrosynthesis.
• Maintained accuracy without loss during the acceleration process.
• Demonstrated improved scalability and industrial applicability of transformer models.

Conclusions:

• Speculative decoding and Speculative Beam Search significantly enhance the efficiency of transformer models for chemical synthesis planning.
• These advancements increase the potential for widespread adoption of advanced AI models in synthetic chemistry.