Characterising biological and environmental samples at the molecular level relies on tandem mass spectrometry (MS/MS), yet interpreting the resulting spectra remains a central challenge: of the roughly 1.7 billion spectra in public repositories, only about 8% can be annotated, because established methods annotate a spectrum only by matching it against a library of previously characterised molecules. The team previously introduced DreaMS (Nature Biotechnology, 2025), a transformer pre-trained self-supervised on hundreds of millions of unannotated spectra, which reaches state-of-the-art annotation accuracy but is limited to matching a spectrum against structures that already exist in a library. Here, the researchers will develop DreaMS-Mol, a foundation model that represents spectra and molecular structures in one shared space, enabling both retrieval of known molecules and de novo generation of new candidate structures for molecules that appear in no library. Trained on reference libraries an order of magnitude larger than before, DreaMS-Mol would extend annotation beyond the small, known fraction of chemistry, reaching the molecules behind the many spectra that remain unannotated across public repositories and biological samples.
Principal Investigator, Company and Country
Tomáš Pluskal, Czech Academy of Sciences, Czechia