NCIT P302 Treatment Indications

NCIT:P302 = Accepted_Therapeutic_Use_For:

"A property representing a disease or condition for which this drug is an accepted treatment. Used in the Drug, Food, Chemical or Biomedical Material branch." (range xsd:string)

It links a drug NCIT concept to a free-text description of the disease / condition it is an accepted treatment for — e.g. Midostaurin (NCIT:C1872) → "acute myeloid leukemia (AML) who are FLT3 mutation-positive (FLT3+)". 796 NCIT drug/chemical concepts carry it.

This is the indication, not the molecular target. The coded molecular-target relation in NCIT is a separate predicate (NCIT:A7 Has_Target, e.g. Midostaurin → FLT3 / PKC / VEGFR2); it is not ingested here but is a natural follow-on for the same generic source.

Two uses

1. Evidence for treatments. P302 is a curated NCI assertion (no PMID), so it is ingested as a structured-database reference source (like ORPHA:, CGGV:, ICEES:) rather than the PMID-snippet path. Each drug becomes one references_cache/NCIT_<Cxxxx>.md cache file holding a unified edge table:

| ID | LABEL | PRED | TARGET_ID | TARGET_LABEL | METADATA |

A curator cites the assertion directly:

yaml treatments: - name: Midostaurin treatment_term: preferred_term: Pharmacotherapy term: { id: NCIT:C15986, label: Pharmacotherapy } therapeutic_agent: - preferred_term: midostaurin term: { id: NCIT:C1872, label: Midostaurin } evidence: - reference: NCIT:C1872 supports: SUPPORT evidence_source: OTHER snippet: "Midostaurin | Accepted_Therapeutic_Use_For | - | - | acute myeloid leukemia (AML) who are FLT3 mutation-positive (FLT3+)" explanation: NCI Thesaurus asserts accepted therapeutic use for FLT3+ AML.

The reference validator substring-checks the snippet against the cache file exactly as it does for ORPHA/ClinGen rows (verified: a wrong snippet fails, the exact row passes).

2. Completeness checking. The P302 indication string is left verbatim as the evidence snippet — it is not parsed. scripts/ncit_p302_audit.py therefore does a pure identifier join on the drug (NCIT concept id, or its ChEBI cross-reference via NCIT P368) against dismech therapeutic_agent ids, flagging each P302 drug PRESENT_WITH_EVIDENCE / PRESENT_NO_EVIDENCE / ABSENT.

How it is built

The generic, manifest-driven OntologyEdgeSource (src/dismech/structured_sources/ontology_edges.py) selects predicate edges out of the OAK-managed sqlite:obo:ncit database — the ontology .db is never committed; only the 796 selectively generated cache files are. The NCIT selection is pinned in data/ncit-edges/MANIFEST.yaml (ontology version 26.02d, predicate NCIT:P302).

just ncit-edges-refresh                 # ensure OAK NCIT db present, check version
just ncit-edges-rebuild                 # regenerate all references_cache/NCIT_*.md
just ncit-edges-rebuild --id NCIT:C1872 # one drug
just ncit-p302-audit --format summary             # per-drug coverage digest
just ncit-p302-audit --format summary --by-class  # rolled up by NCIT is-a class
just ncit-p302-audit --format tsv --out output/ncit_p302_audit.tsv

Audit snapshot (NCIT 26.02d)

The join is on drug identity only (no parsing of the indication text), so these counts are exact, not heuristic. A committed snapshot of all 796 rows lives at projects/NCIT_TREATMENT_INDICATIONS/audit.tsv.

• P302 drug assertions: 796
• PRESENT_WITH_EVIDENCE: 29 — curated in three batches (oncology targeted agents and biologics plus e.g. Paricalcitol/CKD-MBD, Dacarbazine/Hodgkin, BCG/bladder, Siltuximab/Castleman, Olaparib/BRCA ovarian, Enzalutamide/prostate, Sacrosidase/sucrase-isomaltase deficiency, Omega-3/hyperlipidemia).
• PRESENT_NO_EVIDENCE: 147 — drug already used as a dismech therapeutic_agent, P302 evidence not yet cited. Immediately actionable: drop reference: NCIT:<drug> + the verbatim P302 snippet onto the matching treatment record(s). Curator must confirm the indication string matches the disorder — e.g. Arsenic Trioxide's string says "acute myelocytic leukemia", not promyelocytic, so it does not support the APL entry; roughly a third of reviewed candidates are such mismatches and are skipped. Two recurring edit hazards: some treatments place evidence: before treatment_term (so append into the existing block, do not add a second evidence: key), and some are combination records or carry nested evidence: under target_mechanisms.
• ABSENT: 620 — drug carries an accepted-use assertion but is not a therapeutic_agent in any disorder (treatment/disorder curation leads).

Rolled up by NCIT is-a drug class (--by-class)

The flat list is more actionable rolled up by the NCIT rdfs:subClassOf drug class (transitive named ancestors; anonymous OWL restrictions and universal roots dropped). present = used as a dismech therapeutic_agent. This turns "796 drugs" into class-level coverage and aligns mechanism classes to existing modules via an advisory module_hint. Full table: projects/NCIT_TREATMENT_INDICATIONS/class_audit.tsv. Examples:

87 classes with ≥3 drugs have zero dismech coverage (e.g. radiopharmaceuticals, 5-HT3/serotonin antagonists, bisphosphonates, anthracyclines) — class-level gaps. The is-a rollup is a completeness/grouping lens only; is-a is not added to the evidence cache files (it is classification metadata, not a therapeutic-use assertion).

Data-quality caveat: P302 itself can be wrong

P302 is curated NCI editorial text and occasionally carries upstream errors. A confirmed example in NCIT 26.02d: Ramucirumab (NCIT:C70792, an anti-VEGFR2 antibody for gastric/colorectal/HCC) carries a second P302 value "multicentric Castleman's disease (MCD)" — clinically wrong (MCD is treated with the anti-IL-6 antibody siltuximab, NCIT:C61084). The cache faithfully records both rows verbatim; the error is in NCIT, not the extraction. dismech therefore cites only ramucirumab's gastric row and does not propagate the Castleman's row onto Castleman_Disease.yaml. This is why the curator confirmation step (does the indication string actually match the disorder?) is mandatory and not skippable: it guards against over-broad indications and genuine upstream data errors, not just NER mismatches.

Goal / "be complete"

• [x] Ingest all 796 P302 assertions as citable NCIT: cache files.
• [x] Generic OntologyEdgeSource + manifest + CLI + justfile.
• [x] Identifier-join completeness audit (drug id / ChEBI xref, no NER).
• [~] Add the verbatim P302 evidence to the PRESENT_NO_EVIDENCE drugs' existing treatment records (29 done, 147 remaining; confirm each indication string matches the disorder before citing). Deferred trickier structures (nested-evidence/combination records): Tremelimumab, Ganciclovir, Methylene Blue, Idursulfase, Penicillamine, Adapalene, Calcium Carbonate.
• [ ] Triage the 620 ABSENT drugs against the priority dashboard for new treatment / disorder curation, prioritising by is-a class gaps.
• [ ] Use module_hint classes to suggest conforms_to / therapeutic_modality for the matching treatments.
• [ ] Follow-on: add the coded NCIT:A7 Has_Target predicate to the manifest so drug→molecular-target edges back target-based modules.