Skip to content

MONDO Prioritizer

No longer the curation queue

The outstanding curation queue is now stubs/ — one YAML file per disease, edited by pull request. The prioritizer's ranked output is kept as a browsable pool for discovering MONDO concepts to nominate into that queue, but it is no longer the answer to "what should I curate next". Issue #8969 records why: every cheap ontology feature the score can use (child count, synonym count, aggregator tags) correlates with being a grouping rather than with being worth curating, so the head of the ranking filled with umbrella terms no matter how the weights were tuned.

The MONDO prioritizer builds a candidate pool from MONDO disease rows rather than from ad hoc checklist issues. It scores candidate diseases against local DisMech coverage, then applies a small set of explicit specificity heuristics to suggest whether a term should be curated as a root, lumped into a parent, or dropped.

The goal is practical queue construction:

  • reward missing diseases that already have useful MONDO metadata
  • reward external evidence signals when they are available in the input rows
  • penalize terms that are already curated, obsolete, or likely bad roots
  • make every score component inspectable and easy to retune

CLI

uv run dismech-mondo-prioritize \
  examples/mondo_prioritizer_candidates.tsv \
  --format table

Example with a custom config and TSV output:

uv run dismech-mondo-prioritize \
  examples/mondo_prioritizer_candidates.tsv \
  --config examples/mondo_prioritizer_config.yaml \
  --format tsv \
  --output /tmp/mondo_priority.tsv

Generate the static website dashboard from the same scoring logic:

just gen-priority-dashboard

This writes dashboard/priority.html and dashboard/priority.json, and patches dashboard/index.html with a link when that dashboard index already exists.

For a local-only run across all MONDO disease descendants currently missing from kb/disorders/, use:

just gen-priority-dashboard-all-mondo

That flow first exports candidate rows from the local MONDO sqlite database at ~/.data/oaklib/mondo.db, then writes the resulting dashboard under tmp/priority-dashboard-all-mondo/. The tmp/ tree is gitignored so these large artifacts stay out of GitHub by default.

Candidate Export Sidecar

scripts/export_mondo_priority_candidates.py --kb-dir kb/disorders drops already-curated roots before writing the TSV, so the exported rows are the remaining queue only. That leaves the export unable to answer "how many disease terms are already covered?" — a dropped row has no trace in the file, and a TSV has nowhere to record one (a comment line would break csv.DictReader).

The exporter therefore writes a sidecar next to every export, named for it — candidates.tsv → candidates.meta.json:

{
  "disease_root_id": "MONDO:0000001",
  "kb_dir": "kb/disorders",
  "total_descendants": 27438,
  "excluded_curated": 2932,
  "exported_rows": 24506
}

The dashboard reads excluded_curated and counts it in both halves of the coverage fraction: those terms are curated, and they were candidates. Without it the headline reported already_curated: 0 / coverage_percent: 0.0 no matter how much of MONDO the KB covered (issue #7425). A candidate file with no sidecar — a hand-built TSV, or one exported without --kb-dir — is read as "nothing was excluded", which is the correct reading for both.

Expected Input

The prioritizer accepts tsv, csv, json, or jsonl candidate rows. The only required fields are:

  • mondo_id
  • label

Optional fields become weighted signals when present:

  • definition
  • synonyms
  • parents
  • xrefs
  • child_count
  • clingen_definitive_count
  • clingen_strong_count
  • subset_match_count
  • orphanet_match_count
  • is_obsolete

Field aliases such as id, name, disease_mondo, and term_label are also accepted. Multi-valued columns can use ; or |.

Scoring Dimensions

Default weights live in conf/mondo_prioritizer.yaml.

By default the scorer uses:

  • missing_from_dismech: positive base reward for uncurated roots
  • has_definition: reward for a candidate that already has a MONDO definition
  • synonym_count, xref_count: small metadata completeness rewards
  • clingen_definitive_count, clingen_strong_count: optional evidence boosts when those counts are supplied in the MONDO candidate export
  • subset_match_count, orphanet_match_count: optional prioritization boosts when the upstream pipeline already tagged the disease as belonging to an important slice
  • already_curated_penalty, obsolete_penalty: strong penalties for terms we should not queue

Count-based features are capped so a row with very large synonym or xref counts does not dominate the queue.

Specificity Heuristics

The prioritizer also emits a specificity_bucket and recommended_action.

Local MONDO coverage includes each entry's primary disease_term, its declared has_subtypes terms, and mappings.mondo_mappings whose predicate is skos:exactMatch or skos:narrowMatch. A skos:broadMatch, skos:closeMatch, or skos:relatedMatch remains useful as a cross-reference, but does not mean the mapped MONDO concept has itself been curated and therefore does not retire that concept from the queue.

Default buckets are:

  • already_curated
  • obsolete
  • grouping_term
  • subtype_series
  • broad_parent
  • over_specific_leaf
  • root_candidate

The default heuristic rules are intentionally conservative:

  • grouping_term: regex match on high-confidence phrases such as susceptibility, predisposition, or obsolete
  • subtype_series: label matches a configurable subtype suffix pattern such as long QT syndrome 1 or Noonan syndrome 1, and the same stem appears multiple times
  • broad_parent: candidate has many children and is not already classified as a subtype series or grouping term
  • over_specific_leaf: leaf term with a long conjunction-heavy label, which is a useful review flag for phenotype-enumerating leaves

This does not try to solve lump-vs-split perfectly. It surfaces explicit review recommendations such as CURATE_ROOT_WITH_SUBTYPES, LUMP_INTO_PARENT, and REVIEW_AGAINST_PARENT so the queue is defensible and inspectable.

Example Interpretation

With the bundled example candidates:

  • long QT syndrome scores as a missing broad parent and is recommended as CURATE_ROOT_WITH_SUBTYPES
  • long QT syndrome 1 and long QT syndrome 3 score as LUMP_INTO_PARENT
  • autism, susceptibility to, X-linked 3 is flagged as a grouping_term
  • Noonan syndrome is heavily penalized as ALREADY_CURATED
  • obsolete unclassified cardiomyopathy is dropped as obsolete

Why This Helps

Issue-driven checklists are useful prompts, but they mix roots, subtype series, grouping terms, and questionable leaves. A MONDO-first prioritizer gives us a reusable way to:

  • start from MONDO terms and metadata
  • layer in optional external evidence counts
  • compare against current DisMech coverage
  • generate a tuneable queue with transparent reasons

That makes it easier to review prioritization choices and to adjust weights as our curation objectives change.