Glycan neighbourhoods

The distance. Two structures are joined when d = (d_res + d_lnk + d_seq) / 3 falls below the threshold. Equal thirds, fixed.

d_res, d_lnk are Bray-Curtis dissimilarities over residue and linkage counts. Counts, not presence: Bray-Curtis is a frequency measure, and that is the point: set Jaccard scores Glc-Glc-Glc-Gal identical to Gal-Gal-Glc-Glc. This is a deliberate difference from the other four Toolkit tools, which all deduplicate because they answer a different question (“does this structure contain this feature at all?”). The reducing-end marker (b1-OH / a1-OME) counts as a linkage here, as it does in Relationship Explorer.

d_seq is a token-level edit distance over a canonical unit sequence. Two corrections over the reference implementation: branch order is canonicalised first, because [A][B]X and [B][A]X are the same molecule and used to score as different; and one residue or one linkage is one edit unit, because a raw character edit distance made the cost depend on how long a residue's name happens to be.

What this metric is, honestly. The combination (not the weights) was selected from seven candidates by which correlated best with taxonomy. That correlation was r = 0.146, i.e. it explains roughly 2% of the variance in taxonomic distance. So “residual” or “convergent” is very nearly the whole dataset, not a narrow subset. Treat a crossing-phyla group as a prompt to look, not as a result.

The layout is computed from the links. A force simulation is run per connected group when the data is built, so structures that are joined really are drawn near each other and a cluster on screen is a cluster in the data. Distance between two unconnected groups still means nothing; only the links do.