CACTUS: Chemistry Agent Connecting Tool Usage to Science.

Area of Science:

• Computational chemistry
• Artificial intelligence in scientific discovery

Background:

• Large language models (LLMs) show promise but struggle with domain-specific knowledge and tools.
• Accessing and reasoning over specialized chemical information is crucial for molecular discovery.

Purpose of the Study:

• To introduce CACTUS, an LLM-based agent designed to integrate cheminformatics tools for enhanced chemical reasoning.
• To evaluate CACTUS's performance across various open-source LLMs on chemistry-related tasks.

Main Methods:

• Developed CACTUS, an agent that connects LLMs with cheminformatics tools like RDKit.
• Evaluated CACTUS using diverse LLMs (Gemma-7b, Falcon-7b, MPT-7b, Llama3-8b, Mistral-7b) on a large chemistry question benchmark.
• Investigated the impact of domain-specific prompting and hardware configurations.

Main Results:

• CACTUS significantly outperformed baseline LLMs in chemistry problem-solving.
• Gemma-7b, Mistral-7b, and Llama3-8b models showed the highest accuracy with CACTUS.
• Prompt engineering and hardware choices impact performance, with potential for smaller models on consumer hardware.

Conclusions:

• CACTUS effectively bridges LLMs and cheminformatics tools for advanced chemical reasoning.
• The agent facilitates molecular property prediction, similarity searching, and drug-likeness assessment.
• CACTUS empowers researchers by combining LLM capabilities with practical chemical tools.