3115 compounds projected from their record text into two dimensions
(PACMAP over BAAI/bge-large-en-v1.5).
Proximity means described similarly: same pathway,
family, activity, producing organism, where it was found. Not structurally
similar.
Pathways separate without the projection being told them: 84%
of a compound's ten nearest neighbours share its filing pathway, against a 20% baseline.
Measured on this corpus, not inherited.
That is a property of the map you are looking at — the underlying embedding scores 83%, and
PaCMAP tightens neighbourhoods by construction.
Proximity means described similarly — same pathway,
family, activity and producer — not structurally similar. These are text embeddings of
what the corpus says about a compound; SMILES and InChI are deliberately excluded, because a
sentence model reads them as noise. A chemical-similarity map is a different artifact.
Hover a point for its name; click to open the record.
The scatter is a visual aid, and mouse-driven: its points are
canvas pixels, not focusable links, so it is not navigable by keyboard or screen reader.
Nothing is only available here — every compound has a record page reachable from
Browse, and the per-pathway pages list the same records
in text. Said plainly rather than implied, because a scatter that is the sole route to
some records would be an accessibility failure rather than a limitation.