DEEPNET-LLM: Highly accurate structure prediction of unknown metabolites from mass spectra via their molecular neighborhoods

DEEPNET-LLM predicts the correct de novo structure from untargeted mass spectra for 84.7% of spectra (2,249) across 900 authentic, unreleased metabolite standards.

• Why it matters: Most untargeted metabolomics features remain unannotated “dark matter,” and fragmentation spectra are non-unique with an effectively unbounded candidate space, limiting top-ranked de novo structure recovery.
• What they did: The method uses Molecular Community Networking to restrict each query spectrum to a local neighborhood in the global metabolome, then propagates structural information with large language model reasoning.
• The result: It maintains top-1 accuracy independent of precursor mass and comparable across ion polarity, adducts, and in-source fragmentation, reaching 85.9% on MassSpecGym (excluding stereochemistry) and enabling a mouse gut bile-acid tryptamine amidate discovery confirmed by a synthetic standard.

AI summary of the abstract · may contain errors · not medical advice · read the paper.