The problem. Untargeted metabolomics detects thousands of small-molecule features, and most can’t be identified by matching to a reference spectrum — the databases are too incomplete. You’re left with a vast list of unknowns, and no way to organise them or reason about which ones relate to which.

The idea. Molecular networking clusters spectra by similarity: molecules with related structures fragment similarly, so a network of spectral edges groups a known compound with its unknown analogues, letting identity propagate across the graph. Feature-based networking adds the crucial upgrade — linking that network to quantitative feature-detection output (from tools like XCMS, next), so you get abundance across samples, not just connectivity.

Why it matters. This extends the metabolomics arc (sacurine, day 8) from “measure the metabolome” to “make sense of its unknowns.” Conceptually it’s a graph/community approach — the same relational thinking as Louvain/Leiden clustering, applied to spectra instead of cells. It’s also a nice example of coupling qualitative structure (the network) with quantitative data (the features), which is exactly what makes a method usable for real comparisons.

Verdict. Foundational for untargeted metabolomics interpretation and the GNPS ecosystem. Read it for the networking idea — organise unknowns by relationship — and as the analysis counterpart to the peak-picking that feeds it.