Small molecule drugs account for more than half of all FDA-approved medicines and include some of the field’s most celebrated cures.
At Edison, our mission is to build AI platforms across the full R&D life cycle to bring superintelligence to medicine. A critical step toward that goal is accelerating the identification and characterization of new therapeutic candidates, including determining a molecule’s structure.
While nuclear magnetic resonance (NMR) spectroscopy serves as the gold standard for structure elucidation, interpreting NMR spectra remains labor-intensive and requires specialized expertise to navigate the astronomical space of possible molecular structures.
Today, we are introducing ChefNMR2, a novel AI system for rapid, automated de novo structure elucidation from simple one-dimensional (1D) NMR spectra to substantially expedite the advancement of small molecule drugs.
In collaboration with Princeton’s Ellen Zhong, Assistant Professor of Computer Science, and Mo Seyedsayamdost, Professor of Chemistry, our work brings together two complementary advances at the forefront of modern AI: specialized atomic generative models and general-purpose large language models (LLMs).
- ChefNMR2: a state-of-the-art molecular generative model that uses an atomic flow-matching architecture trained on 2.52 million molecules with experimental and simulated NMR spectra
- ChefNMR2-Agent: an agentic workflow that augments ChefNMR2’s molecular predictions with LLM-based chemical reasoning
Going complex
Despite their emerging chemical reasoning capabilities, LLMs remain limited in accuracy and reliability, particularly for complex molecules.
To test ChefNMR2’s ability to address this gap, we evaluated nine well-known natural products spanning structural classes. ChefNMR2 determined the correct planar structure for all nine, including vancomycin aglycone, rapamycin and taxol, the largest and most complex examples in this set.
Compared with six frontier LLMs, ChefNMR2 achieves substantially higher accuracy, well-calibrated confidence estimates, and up to 2,170-fold faster inference and 360-fold lower cost.
Notably, across five diverse benchmarks comprising 2,031 molecules, ChefNMR2 ranked the correct structure first in 82.4% of cases, compared with 51.9% for standalone Claude Opus 5, the strongest general-purpose LLM tested.
Elucidation of unreported natural products
ChefNMR2 was integrated into our AI scientist, Kosmos, to elucidate the structures of four unpublished natural products isolated by the Seyedsayamdost lab. Our system matched human-chemist accuracy in a fraction of the time by using only 1D NMR spectra and molecular formulae, demonstrating how agents can effectively improve the predictions of specialized models.
Additionally, by integrating molecular generation with chemical reasoning, chemical-shift prediction, and literature search, the agent can resolve structures that elude both standalone generative models and LLMs, including a natural product with an unprecedented prenyl substitution.
ChefNMR2 is now available in Kosmos for our partners to accelerate small molecule drug discovery.
Project page: https://chefnmr.cs.princeton.edu
Web app & MCP tool: https://huggingface.co/spaces/ziyux123/chefnmr2