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

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Organic synthesis

Background:

• Generative AI models show promise for small molecule drug discovery.
• A key limitation is the generation of molecules with complex or impractical synthetic routes.
• This hinders the progression of AI-discovered compounds into further development.

Purpose of the Study:

• To develop a novel AI approach for generating synthesizable molecules and their complete synthetic pathways.
• To fine-tune Meta's Llama3 Large Language Models (LLMs) for robust organic synthesis planning.
• To create SynLlama, an AI tool that addresses the synthesizability challenge in generative drug discovery.

Main Methods:

• Fine-tuning Meta's Llama3 Large Language Models (LLMs) to create SynLlama.
• Training SynLlama on organic reaction templates and accessible building blocks.
• Evaluating SynLlama's performance in bottom-up synthesis, analog generation, and hit expansion.

Main Results:

• SynLlama generates full synthetic pathways using accessible building blocks and reaction templates.
• The model explores a vast synthesizable chemical space with less data than state-of-the-art methods.
• SynLlama demonstrates effective generalization to unseen building blocks, expanding its reconstruction capabilities.
• Successful application in synthesis planning for analog molecules and hit expansion in a pharmaceutical context.

Conclusions:

• SynLlama offers a valuable tool for medicinal chemists, overcoming synthesis barriers in generative drug discovery.
• The model significantly enhances the practicality and efficiency of AI-driven small molecule design.
• SynLlama's ability to generalize expands the accessible synthesizable chemical space for drug development.