NISPO: Open-source AI-built tool generates IUPAC names with 98.1% accuracy
An AI agent trained on 2.68M molecules now names 103M compounds with near-perfect accuracy.
A team led by Nicholas Runcie at Oxford has released NISPO, an open-source RDKit-based Python package that generates systematic IUPAC names for chemical structures. What makes this tool novel is its development process: an AI agent powered by OpenAI's Codex (using the GPT-5.5 model) iteratively improved the naming algorithm through a self-improvement loop. Correctness was validated via round-trip parsing—if the open-source OPSIN tool could parse the generated name back to the original structure, it was considered correct. The agent trained on 2.68 million molecules from the SureChEMBL patent database, continuously refining its approach.
On a held-out test set of 103 million molecules from PubChem, NISPO achieved 98.1% round-trip accuracy, a remarkable feat for an automatically-generated tool. The package is freely available on GitHub, making it accessible to researchers, patent analysts, and drug discovery teams who need reliable IUPAC naming without proprietary software. This work demonstrates how AI-driven iterative development can produce high-quality scientific software with minimal human intervention.
- Built using an agentic self-improvement loop with OpenAI Codex (GPT-5.5 model) to optimize IUPAC naming rules.
- Achieved 98.1% round-trip accuracy on 103 million PubChem molecules after training on 2.68 million SureChEMBL compounds.
- Fully open-source Python package built on RDKit, freely available for academic and commercial use.
Why It Matters
Free, high-accuracy IUPAC name generator democratizes chemical communication for researchers and patent analysts alike.