Chemical structure map
This view places 3115 records using chemical structure only: chiral Morgan count fingerprints and Tanimoto distance, projected to two dimensions with UMAP. Labels, filing pathways, producers, and curation fields can color or filter the view, but they do not affect position.
Read local neighborhoods, not the axes or global spacing. UMAP preserves nearby structure better than map-wide distance. Current full-corpus trustworthiness@10 is 0.975 and 2D neighbor overlap@10 is 0.537. Multi-fragment structures are retained exactly as their records define them and are visibly marked.
Method and provenance
Model morgan-count-chiral-r2_0.90-r4_0.10+tanimoto+umap-precomputed-n15-d0.05-c2-rs42. The generator parses stored SMILES
first and falls back to standard InChI when RDKit rejects the stored
SMILES. This build used that fallback for
0 records and retained
19 multi-fragment records.
Of 129 detected stereoisomer pairs,
128
have nonzero model distance.
The remaining
1 sit at distance zero
and therefore on one point: a Morgan fingerprint encodes tetrahedral
chirality through its atom invariants and does not encode sulfoxide
stereochemistry, so sulfoxide epimers hash identically at every radius.
Those records are distinct and correct — their InChIKeys differ — and it
is this map's distance metric that cannot separate them.
Distance is 90% radius-2 and 10% radius-4 chiral Morgan count-fingerprint
Tanimoto distance. Coordinates are generated with
UMAP(metric=precomputed, n_neighbors=15, min_dist=0.05,
random_state=42). A corpus or model-version change can move every
point.