#37 · Proteomics

Transformer and foundation models are reshaping tandem mass spectrometry interpretation

The dominant thread is learned models replacing or augmenting hand-engineered steps in tandem mass spectrometry workflows. LipiDetective applies a transformer trained on lipid standard and database spectra to annotate lipid species independent of acquisition setup. yHydra embeds spectra and peptides jointly in a foundation model to support open and error-tolerant searches, reporting up to 25% more identifications. pFind+ adds deep learning rescoring and open search for PTM discovery, and the acetylation study fine-tunes AHLF transfer learning on roughly 500,000 peptide-spectrum matches from nine datasets, indicating modest data needs for acetylation and ubiquitination PTM identification. GLACIER is more distinct, recasting MS/MS prediction as single-stage transformer fragment detection on molecular graphs. aiSysMet contributes a cloud-executable platform for LC-MS metabolomics processing and multi-omics integration. What is new across the set is the emphasis on transferable representations, exemplified by the yHydra foundation model and AHLF transfer learning.