#2 · ML / AI methods

Generative models map mass spectra directly to candidate molecular structures

The dominant thread is generative modeling that conditions molecular output on experimental mass spectrometry observables. FRIGID, a diffusion language model, generates molecular structures from mass spectra via intermediate fingerprints and formulae, and adds inference-time fragment remasking. MetGenX, a structure-informed encoder-decoder network, generates metabolite structures directly from MS2 spectra and reaches 55.9% top-1 accuracy on NIST spectra. π-MSNet is a complementary data resource rather than a generative method: it provides a continuously reprocessed, AI-ready collection of 737M PSMs from 45,583 LC-MS/MS runs for benchmarking MS2, RT, CCS, and de novo tasks. What is new is direct structure generation from spectra, with FRIGID explicitly scaling both training and inference-time refinement. The items are heterogeneous in domain, since π-MSNet targets proteomics, so the thread is strongest for the two metabolite-focused generators.