The problem. In metabolomics MS/MS, co-eluting compounds get fragmented together, so a single spectrum is often a blend of ions from several metabolites. Trying to match that chimeric spectrum to a single database entry fails — the mixture looks like nothing in the reference, and real compounds go unidentified.

The idea. DecoID computationally deconvolves these mixed MS/MS spectra with help from a spectral database: it models an observed spectrum as a combination of database components plus residual, and solves for which known compounds (and how much of each) best explain the mixture. Pulling the blend apart recovers identifications that a naive one-to-one match would miss.

Why it matters. This is the same problem, one modality over, that I keep meeting: DIA-NN deconvolves co-eluting peptides, EmptyDrops separates ambient from real, molecular networking relates unknowns — all variations on “the measurement is a mixture; recover the components.” Seeing the pattern recur across proteomics, single-cell, and metabolomics is the real lesson of this toolchain day: interference is universal, and good methods model it.

Verdict. A solid, focused methods contribution that meaningfully raises identification rates in untargeted metabolomics. Read it as the deconvolution capstone to the metabolomics thread — and as one more instance of modelling mixtures instead of pretending they’re pure.