Problem solved

Chemists and researchers often spend considerable time on routine tasks like computing molecular properties, searching databases for compounds, planning syntheses, and ordering chemicals. General LLMs lack access to accurate, up-to-date chemical data and cannot execute specialized cheminformatics operations. ChemCrow solves this by bridging LLMs with a suite of chemistry tools, allowing users to perform these tasks through natural language instructions, thus automating and accelerating chemical research and development workflows.

Use cases

• Calculating molecular similarity (e.g., Tanimoto similarity between two compounds using ECFP fingerprints)
• Predicting toxicity of a given chemical structure
• Planning a synthetic route for a target molecule
• Purchasing a specific chemical from an online catalog like A2B Chem
• Retrieving detailed compound information from public databases such as PubChem
• Evaluating drug-likeness or other ADMET properties of drug candidates

Audience

• Researchers
• Data Scientists
• ML Engineers

Pros

• Integrates multiple specialized chemistry tools into a single conversational interface
• Automates routine cheminformatics and procurement tasks
• Leverages the advanced reasoning capabilities of LLMs like GPT-4
• Potential to accelerate drug discovery and materials research
• Open-source research project that invites community validation and improvement

Cons

• The LLM evaluator cannot reliably distinguish correct from incorrect outputs, posing safety risks
• Inaccuracies in practical tasks (e.g., mispriced chemical, inappropriate toxicity classification) undermine trust
• Citation practices and tool integration details may be flawed, as evidenced by peer review comments