Although metabolism has been a focus of relatively intense study for more than a century, a huge number of its molecular byproducts remain unknown. Even when small-molecule metabolites are detected by technologies such as mass spectrometry, they often cannot be identified or structurally described. Large language models (LLMs), which can be trained on the chemical structures of small molecules via formats that represent these structures as short strings of text, could help solve this problem. Such models are already widely and productively applied to predict the structures of macromolecules like proteins, discern their functions and design variants with specific functions based on the evolutionary forces that have shaped the original molecule. The application of LLMs to small molecules, however, has been restricted primarily to generating structures of potential utility as drugs. Researchers led by Ludwig Princeton’s Hantao Qiang and Michael Skinnider reported in a January paper in Nature their development of an LLM named DeepMet to explore the vast space of unknown metabolites—often called the “dark matter of the metabolome”. DeepMet learns from the structures of known metabolites to anticipate the existence of previously unrecognized metabolites. Its integration with mass spectrometry-based metabolomics data facilitates metabolite discovery and the researchers illustrated its power by applying DeepMet to describe several dozen structurally diverse mammalian metabolites.
Language model-guided anticipation and discovery of mammalian metabolites
Nature, 2026 January 14