Interpreting 200 million mass spectra of small molecules using DreaMS neural network

It is estimated that over 90% of small molecules from nature remain undiscovered, holding opportunities for progress in molecular biology, drug discovery, and environmental research. Meanwhile, public metabolomics repositories such as MassIVE GNPS contain millions of uncharacterized tandem mass spectra corresponding to these undiscovered compounds. Here, we introduce DreaMS—a large language model for interpreting tandem mass spectra of small molecules. Trained in a self-supervised way on spectra from MassIVE GNPS, DreaMS learns rich representations of underlying molecular structures. We use these representations to build the DreaMS Atlas: a molecular network of 200 million spectra that provides a new way to connect and analyze molecules across thousands of distinct scientific studies. We also show that DreaMS can be fine-tuned for virtually any mass spectrum interpretation task. For example, fine-tuning DreaMS to identify fluorinated compounds led to the discovery of a new class of plant fluorinated metabolites (at least 9 new structures) while previously only several relatively simple structures were known.