#40 · Metabolomics

Learned priors and fragment-ion evidence under validation in MS/MS identification

The dominant thread is that MS/MS identification is being evaluated against explicit reference data and constrained by learned or structural priors. The published benchmark of eight de novo peptide sequencing models on three proteomics datasets reports shared errors tied to limited fragment coverage, indicating that model class alone does not resolve sequence ambiguity. The preprints extend this logic beyond peptides: FACET enumerates the complete feasible set of molecular formulas from precursor mass before neural ranking, Carafe uses deep learning fragment-ion intensity prediction to prune the Comet fragment ion index, and DEEPNET-LLM restricts de novo metabolite structure prediction to molecular-community neighborhoods. What is new is treating fragment-ion evidence and the priors placed on it as the variables under validation.