A mass spectrometer works like a molecular shredder: it breaks a substance into charged fragments and weighs each one. From this 'fingerprint,' chemists can tell what was in the sample. But if the molecule is new and its fragments don't match anything known, the instrument just shows a set of peaks, and what lies behind them is a mystery.
The authors of the paper proposed a workaround. They took an existing language model that understands chemical structures as text and trained a neural network to translate spectral peaks directly into that semantic space. Now even an unfamiliar molecule lands near structurally similar compounds. In a strict exam where all molecules were new to the system, it correctly named 42.2% of 256 candidates, like finding the right book in a huge uncatalogued library by a single phrase from it.
This approach could speed up drug discovery or the analysis of complex mixtures. Instead of manually sifting through databases, a chemist gets a hint: 'here's what this looks like'—even when there's no exact match.
Mass spectrometry helped determine that the Murchison meteorite contains more than 70 amino acids, the building blocks of life, many of which are not found on Earth.