references/curation-rules.md
Curation rules
How to choose among candidates, and what to do when the honest answer is "no term".
The decision procedure
Run it per string. Stop at the first step that gives a defensible answer.
- Exact label match in the expected ontology, from its defining ontology. Check its definition, current term detail, organism/context and target-schema constraints before accepting. Multiple exact candidates require disambiguation; a bounded top-hit list is not exhaustive.
- Exact synonym match with an exact synonym annotation. Apply the same context checks and record the primary label. A related/broad/narrow or unscoped synonym is a curation candidate, even when its spelling exactly matches the input.
- Exact match, wrong ontology. Usually a category error in the source column, not a naming
problem —
hepatocytein a tissue field means the column mixes tissue and cell type. Fix the column, do not force a match. - Partial match only. Do not accept silently. Either:
- normalise the input and retry (see below), or
- present the top candidates with their labels and let a human choose, or
- mark it unresolved.
- Nothing. Mark unresolved and say so. An unresolved row is a correct output.
resolve_terms.py implements steps 1–4's search side and labels every hit exact_label,
exact_synonym, scoped/unscoped synonym matches, or partial. The judgement about a match is yours;
the tool will not make it for you.
Normalisations worth retrying
Cheap rewrites that convert a partial into an exact_label, in rough order of yield:
- Separate preparation and donor notes from anatomy, preserving them in other fields. Do not
reduce
Liver - left lobe [FFPE]toliver: the left-lobe qualifier carries anatomy. - Expand lab shorthand:
PBMC→peripheral blood mononuclear cell,WT→ the actual genotype,M/F→male/femaleonly with an explicit source codebook. - Reverse an inverted phrase:
ventricle, left→left ventricle,cortex, kidney→kidney cortex. - Singularise:
hepatocytes→hepatocyte. Check the actual label; labels are not universally singular. - Anglicise or Americanise: ontology labels vary; try both
oesophagusandesophagus. - Strip species prefixes:
human liver→liver(species belongs in a separate NCBITaxon field).
Do not normalise away hyphens, Greek letters, digits, or capitalised gene symbols — CD4-positive
and alpha-beta T cell mean what they say, and normalize_label() deliberately folds only case and
whitespace.
When plain search keeps failing on lab shorthand, run map_terms.py with --ontology set
(see companion-apis.md). ZOOMA can combine curated, lexical and embedding evidence. Inspect underlying provenance;
its v2 wrapper can call non-curated matches ZOOMA_INFERRED_FROM_CURATED. Every HIGH/GOOD
CURIE still goes through validate_terms.py before it is written down.
What "unresolved" should look like
Never invent an ID to fill a cell. An unresolved row carries the original string, an empty ID, and
the reason. Downstream that is a visible gap; an invented UBERON:0002108 is a silent error that
survives review because it looks exactly like a real ID.
If a concept genuinely has no term and the project depends on it, the route is a new-term request to the ontology (GitHub issue on the ontology's tracker, with a definition and a reference), not a locally minted identifier.
Auditing an existing metadata table
The high-yield checks, in order:
- Every ID exists.
validate_terms.py --input table.tsv. - No obsolete IDs. Obsolete terms carry
term_replaced_byoften enough that the fix is straightforward — but apply replacements deliberately, since a replacement can be broader or narrower than the original. - Labels match IDs. Supply the label column. Mismatches are where copy-paste drift and hallucinated IDs surface: the ID is real, the label is real, and they describe different things.
- Right ontology per column.
--expect-ontology. - Right branch per column.
--branch, remembering it does not exclude cell types from anatomy (ontology-registry.md).
--strict turns warnings into failures, which is the right setting for a CI gate. Warnings are
matched_synonym (label is a synonym rather than the primary label), imported_only (no defining copy verified by this lookup), and not_a_class.
Obsolete terms
Obsoletion is not deletion — the ID keeps resolving, and its label is usually prefixed
obsolete_. That prefix is a useful smell in any metadata file:
EFO:0001067 obsolete_parasitic infection -> replaced by MONDO:0005135
Some obsolete terms have no replacement, only a consider annotation or nothing at all. Then the
term must be re-curated by hand; there is no automatic answer.
Cross-ontology mapping
OxO2 is live with compatibility APIs. The bundled scripts do not perform cross-ontology mapping; two source-grounded routes are:
- Term cross-references.
term_detail(curie)["annotation"]["database_cross_reference"]lists cross-reference candidates —UBERON:0002107carriesMESH:D008099,NCIT:C12392,FMA:7197,UMLS:C0023884, and more. - SSSOM mapping sets published by Monarch and the OBO community, when provenance and mapping
predicates (
skos:exactMatchvscloseMatch) matter.
Cross-references are asserted by curators at varying confidence and are not all exactMatch.
Treat a single xref as a lead, not a proof, when the mapping drives analysis rather than display.
from ols_client import term_detail
xrefs = (term_detail("UBERON:0002107") or {}).get("annotation", {}).get(
"database_cross_reference", []
)
Reporting
When you hand back resolved terms, give the ID and the label, and say how each was matched. A
table of bare IDs cannot be reviewed — no reader can tell UBERON:0002107 from UBERON:0002108
by eye, which is precisely why invented IDs survive review.