CACTUS: Chemistry\nAgent Connecting Tool Usage to Science

Large language models\n(LLMs) have shown remarkable potential\nin\nvarious domains but often lack the ability to access and reason over\ndomain-specific knowledge and tools. In this article, we introduce\nChemistry Agent Connecting Tool-Usage to Science (CACTUS), an LLM-based\nagent that integrates existing cheminformatics tools to enable accurate\nand advanced reasoning and problem-solving in chemistry and molecular\ndiscovery. We evaluate the performance of CACTUS using a diverse set\nof open-source LLMs, including Gemma-7b, Falcon-7b, MPT-7b, Llama3-8b,\nand Mistral-7b, on a benchmark of thousands of chemistry questions.\nOur results demonstrate that CACTUS significantly outperforms baseline\nLLMs, with the Gemma-7b, Mistral-7b, and Llama3-8b models achieving\nthe highest accuracy regardless of the prompting strategy used. Moreover,\nwe explore the impact of domain-specific prompting and hardware configurations\non model performance, highlighting the importance of prompt engineering\nand the potential for deploying smaller models on consumer-grade hardware\nwithout a significant loss in accuracy. By combining the cognitive\ncapabilities of open-source LLMs with widely used domain-specific\ntools provided by RDKit, CACTUS can assist researchers in tasks such\nas molecular property prediction, similarity searching, and drug-likeness\nassessment.