Skip to content

Concept Explorer and embeddings

The disease and mechanism embedding browsers now live in the DisMech Concept Explorer, alongside ontology associations. Their old URLs redirect there and preserve ?focus=<disease name>&space=<representation> selections.

Choose an ontology term, inspect its associated diseases, and follow Explore similarities and mechanisms into the map. Disease details link to individual mechanisms; mechanism details link to curated annotations, downstream nodes, module conformance, and the original DisMech entry. Clear term selection returns to the whole similarity space.

Freshness and methods

The downstream repository builds daily from a fresh DisMech main checkout and publishes a validated Pages artifact. Its source SHA, data-build timestamp, coverage, model, and build-status link are visible. A release manifest records input hashes, ontology versions and output checksums. Generation failures retain the last successful snapshot rather than publishing incomplete data.

The new local model2vec representations replace the old OpenAI ada snapshots; cosine values and plot positions are not numerically comparable across those models. Neighbors use full-vector cosine similarity; the map offers separate UMAP, t-SNE, and PCA projections. Drag a box to zoom, use the toolbar to pan/reset/export an image, or select points to filter the concept list and CSV export. The sidebar offers category/parent coloring, All/None filters, labels, and focused-neighborhood zoom. Similarity is a curation lead, not evidence of shared causation.

The ontology score implementation is also explicitly versioned as context-v2.1-max-product: the original exporter was not recoverable. Read the method contract for the formula, propagation predicates, exclusions and limitations. The original audit and architecture records why the browsers were combined.

Development

The UI, source extraction, score export and publication pipeline are maintained in dismech-ont-scores. Curated assertions remain here in kb/. To rebuild the browser, run in its repo:

uv sync --locked
uv run python scripts/build_site.py --source /path/to/dismech
uv run python scripts/build_site.py --validate
uv run python -m http.server 8000 --directory dist

No paid API key is needed. The first run downloads the pinned local model and ontology snapshots. Vectors are reused by model/revision/text hash, while the current entity inventory is rebuilt independently, so edits, renames, and removals cannot leave stale entities in the published map.

src/dismech/embed.py, the just embed-* recipes, and committed embedding JS remain available as legacy research tools/artifacts; they no longer supply the public browser. In particular, legacy ordinary indexing does not refresh an existing nonempty collection, and its export enumerates the vector cache; do not use it as a fresh current-KB export. The independent dismech.node_embeddings evaluation/suggestion CLI remains available for research.