VERDANT integration (Project 11)¶

VERDANT builds AI-generated SKOS mappings (exactMatch, closeMatch, relatedMatch) from vocabulary terms to ontologies and knowledge bases, with an expert pre-check loop and a federated SPARQL endpoint.

This page sketches how data could flow between the two projects. Nothing here has been agreed with the VERDANT team.

Data flow¶

flowchart LR
  A[MiCoKnowGraph<br/>unmapped labels TSV] --> B[VERDANT<br/>embeddings + FAISS]
  B --> C[SKOS + confidence<br/>or SSSOM]
  C -->|scripts/skos_to_sssom.py| D[mappings/verdant/candidates/]
  D -->|PR review| E[mappings/curated/]
  D -->|PR review| F[mappings/verdant/rejected/<br/>predicate_modifier: Not]
  F -->|feedback / training data| B
  E -->|make mappings| G[(KG: skos:exactMatch triples)]

1. MiCoKnowGraph → VERDANT¶

Proposed format for the terms we would send (no script produces this yet):

column	example
`term_id`	`micokg:term/habitat/sewage`
`label`	`sewage`
`context_slot`	`env_medium`
`source_kg`	`infores:gutmdisorder`
`suggested_target_ontology`	`ENVO`

Sources: - free-text habitat, sample-type, host and disease labels from existing microbiome KGs; - labels found when real data is loaded.

The three directories under mappings/ are curated/ (reviewed, loaded into the KG), verdant/candidates/ (unreviewed VERDANT output) and verdant/rejected/.

2. VERDANT → MiCoKnowGraph¶

Preferred: SSSOM TSV. Also accepted: Turtle with plain SKOS triples, or reified owl:Axiom mappings with sssom:confidence. See tests/fixtures/verdant-sample.ttl.

python scripts/skos_to_sssom.py verdant-output.ttl \
  -o mappings/verdant/candidates/habitat.sssom.tsv \
  --mapping-set-id https://w3id.org/micokg/mappings/verdant/habitat \
  --tool-version 0.3.0

Every candidate row gets: - mapping_justification: semapv:SemanticSimilarityThresholdMatching - mapping_tool: VERDANT - the tool version and a confidence score - no reviewer_id

3. Review¶

Reviewers go through candidates in a pull request:

Decision	Action
Accept	Move the row to `mappings/curated/<topic>.sssom.tsv`, put your ORCID in `reviewer_id`, keep the original confidence
Accept with a different predicate	As accept, change `predicate_id` (e.g. exact → close), and explain in `comment`
Reject	Move the row to `mappings/verdant/rejected/`, add a `predicate_modifier` column set to `Not`, and give the reason in `comment`

Rejected rows are published so VERDANT can use them as negative training examples.

4. How mappings are used in the KG¶

make mappings validates SSSOM files and writes plain skos:exactMatch / closeMatch triples. The CLI's own output is reified owl:Axioms, which are hard to query.
Queries expand identifiers with ?disease (skos:exactMatch|^skos:exactMatch)?/rdfs:subClassOf* ?queryDisease. See queries/uc3-host-microbiome-cancer/cq01-*.rq.
Only curated mappings are loaded into the KG used by tests. Candidates are not.

Open points to agree with VERDANT¶

The identifier scheme for free-text labels, i.e. who mints micokg:term/... IDs.
Target ontologies and versions: ENVO, UBERON, NCBITaxon, MONDO, ChEBI, FoodOn?
Confidence calibration: is 0.8 comparable across target ontologies?
Endpoint URL, and whether MiCoKnowGraph example queries should be loaded there (SIB sh:SPARQLExecutable files from make spex).
Update cadence and versioning of mapping sets.