Molecular Community Network links nearly 95% of mass-spec molecules to at least one neighbor

Classical molecular networking draws an edge between two molecules only when the similarity score of their mass spectra clears a preset threshold ^[4]. Relatives that score just under the cutoff land in separate pieces, so one chemical family can be split across the map ^[4]. That matters for unknowns, because identification by network works through known neighbors ^[7].

The Molecular Community Network starts from the whole network. Its algorithm partitions it into communities, groups with many strong links inside and fewer between them, and then keeps the strongest connections so the communities stay joined ^[5]. "This approach allows us to have almost every molecule in the network linked to at least one neighbor," said Vladimir Boginski, who led the work with first author Elizabeth A. Coler and an interdisciplinary team ^[1][2]. "Moreover, these links are typically between molecules from similar molecular families." ^[6] An unknown that shares a community with identified molecules can then be given a predicted identity from them. The field calls this step annotation propagation ^[7].

If Boginski's upper estimate of 90% dark matter held across the roughly 8.4 million spectra in public repositories ^[3], up to about 7.6 million of them would be detectable but unmatched to any known structure ^[12][1]. With nearly 95% of molecules now in communities ^[8], about 420,000 or slightly more would still sit outside one ^[2].

The thing this doesn't tell you is how often a propagated identity is right. Connection gives an unknown what Boginski called "a much wider and richer search space for molecular discovery" ^[8]. The public account of the paper does not include an accuracy rate for predicted identities, or a connectivity figure for threshold networking on the same data.

The bile-acid class is the one direct test in that account. The network made the prediction first, and lab work confirmed it afterward ^[9]. That order is the right design, because the prediction could have failed at the bench.

Boginski puts the payoff in biomarker work. "Biomarker discovery often stalls at the point where a molecule that differentiates sick from healthy is detected by its mass spectrum, but it may not be known what this molecule is," he said ^[10]. "MCN can potentially help convert some of those 'dead ends' into identifiable molecules," he said ^[11]. I think "some" is the accurate word, since the published account describes one confirmed class so far ^[9].