SynLlama: Generating Synthesizable Molecules and Their Analogs with Large Language Models

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Drug Discovery

Background:

• Generative machine learning models offer potential for chemical space exploration.
• A key limitation is the frequent generation of molecules with challenging synthesis routes.
• This hinders practical application in drug discovery and development.

Purpose of the Study:

• To develop a novel generative model for creating synthesizable molecules.
• To fine-tune Large Language Models (LLMs) for chemical synthesis planning.
• To create SynLlama, capable of generating complete synthetic pathways.

Main Methods:

• Fine-tuning Meta's Llama3 Large Language Models (LLMs).
• Developing SynLlama to generate synthetic pathways using accessible building blocks and reaction templates.
• Evaluating performance in forward and bottom-up synthesis planning.

Main Results:

• SynLlama effectively generates full synthetic pathways for novel molecules.
• The model explores a large synthesizable chemical space with reduced data requirements.
• SynLlama demonstrates generalization to unseen, purchasable building blocks.
• Strong performance in synthesis planning compared to state-of-the-art methods.

Conclusions:

• SynLlama offers a powerful tool for exploring synthesizable chemical space.
• The model provides practical synthetic routes, overcoming limitations of previous generative approaches.
• SynLlama is valuable for pharmaceutical applications, including analog synthesis and hit expansion.