Large language model guides exploration of the dark metabolome

Large language models may help researchers characterize previously unknown metabolites. Michael Skinnider and his team have developed DeepMet, a large language model–guided program that can assign a structure to an unknown peak on a mass spectrometry graph (Nature 2026, DOI: 10.1038/s41586-025-09969-x).When researchers use mass spectrometry to conduct a metabolomics study, they don’t actually identify most of the signals in the data, Skinnider says. These unknown data points compose “the dark metabolome,” and while some of them may result from fragmenting known metabolites, some of these unexplained data points may be from entirely uncharacterized metabolites. DeepMet was trained on the chemical structures of known mammalian metabolites. Now, with just a little information about a mass spectrometry data point, like the mass, the algorithm comes up with a likely structure for the molecule that corresponds to that data point. “We're borrowing this idea from the context of drug discovery and pharma or biotech companies, where it's been very successful, and applying it for the first time to mammalian metabolism,” Skinnider says.With DeepMet, from a variety of human and mouse metabolomics datasets, Skinnider and his team found several murine and human metabolites that were not present in existing databases.But DeepMet is designed to find structures that are similar to what it was trained on. Skinnider says this means the model won’t catch a metabolite that has completely different chemistry from that of the training dataset. But if there is a metabolite that’s significantly different from other known mammalian metabolites, it’s likely that its