Augmenting large language models with chemistry tools

Area of Science:

• Artificial Intelligence in Chemistry
• Computational Chemistry
• Drug Discovery and Materials Design

Background:

• Large language models (LLMs) show promise but struggle with specialized scientific domains like chemistry.
• Current LLMs lack access to external knowledge, limiting their utility in scientific research.
• There is a need for AI agents that can effectively handle chemistry-specific tasks.

Purpose of the Study:

• To introduce ChemCrow, an LLM-powered agent designed for chemical applications.
• To enhance LLM capabilities in organic synthesis, drug discovery, and materials design.
• To demonstrate the autonomous execution of chemical tasks by an AI agent.

Main Methods:

• Developed ChemCrow by integrating 18 expert-designed tools with GPT-4 as the core LLM.
• Utilized ChemCrow to autonomously plan and execute chemical syntheses and guide discovery processes.
• Evaluated ChemCrow's performance using both LLM-based and expert assessments.

Main Results:

• ChemCrow successfully planned and executed the synthesis of an insect repellent and three organocatalysts.
• The agent played a role in guiding the discovery of a novel chromophore.
• Evaluation confirmed ChemCrow's effectiveness in automating diverse chemical tasks.

Conclusions:

• ChemCrow significantly augments LLM performance in chemistry, enabling new capabilities.
• The agent automates complex chemical tasks, aiding both expert and non-expert users.
• ChemCrow bridges the gap between experimental and computational chemistry, fostering scientific advancement.