Curation Backlog (Cross-Cutting)

In progress METABACKLOGCOVERAGE
Home / Projects / Curation Backlog (Cross-Cutting)

Curation Backlog (Cross-Cutting)

Overview

Most dismech project files are thematic — a disease domain, a drug class, a mechanism. This one is orthogonal: it collects what is outstanding across the whole KB, so that "what should I curate next?" has an answer that does not depend on picking a theme first.

Four workstreams, ordered by yield per unit of effort rather than by size. The first three are triage work — the candidates already exist and need judgement, not fresh literature search. The fourth is genuine curation volume.

Supporting data lives in projects/CURATION_BACKLOG/; regenerate it from the dashboards rather than hand-editing.

How fast this decays

All four workstreams were rebuilt against main @ 48bd0eb60c (2026-09-03). Treat that date as a short expiry, not a footnote: 61 of workstream 1's candidates were curated in the four days between two rebuilds, and the KB grew from 1,937 to 2,535 disorder entries over three weeks.

The first version of this file was assembled from a tree 9,389 commits behind main. The dashboards were regenerated at the time and came back byte-identical, which was read as confirmation the figures were current. It was not: both the committed and the regenerated dashboards derived from the same stale tree, so the check established self-consistency and nothing about currency. git log -1 origin/main settles it in one command — run that before trusting any count here.

Where the KB stands

Verified against main @ 48bd0eb60c (2026-09-03):

Measure Value
Disorder entries 2,535
…with a MONDO-anchored disease_term 2,499
Distinct curated MONDO IDs (primary, subtype, exact/narrow mapping) 3,857
MONDO human-disease terms with an exact dismech page 2,469 of 23,213 (10.6%)
…covered exactly or by a parent page 48.3%
Average weighted compliance 90.0% (684 violations)
Entries declaring any module conformance 1,001 (39.5%)
Comorbidity entries 26
Mechanism modules 164
Groupings 98

Pathophysiology and phenotype coverage are effectively complete. The gaps are everywhere else.

A prior question: is the candidate genuinely uncovered?

Before curating anything from these lists, check that a curated entry does not already cover the concept under a different MONDO term — usually the target's parent. The lexical preflight cannot see this, which is how issue #10069 was filed against a concept Pituitary_Tumor.yaml had covered for ten days.

uv run python scripts/audit_curation_queue_coverage.py --max-descendants 20

Over the 1,338-stub queue that check finds 3 stubs already curated exactly and 641 with a curated ancestor. Those 641 are mostly not duplicates — they are lump/split rulings the queue is presenting as "curate a new entry" (mevalonic aciduria and hyper-IgD both under Mevalonate_Kinase_Deficiency; sclerosteosis 2 under Sclerosteosis). Curating one as a fresh entry creates a contradiction rather than resolving it. Details and the over-broad-anchor finding are in #10679.

Workstream 1 — MONDO terms already linked from existing pages

397 candidates (552 → 458 → 397 across three rebuilds; 61 curated in the last four days alone). Data: linked-but-uncurated.tsv.

These diseases are already referenced by curated entries — as differentials, comorbidities, or related conditions — but have no entry of their own. Demand is therefore already demonstrated, which makes this the cheapest breadth work in the repo: no need to argue the disease is worth curating, only to curate it.

The churn is the argument for re-deriving before picking. Myocarditis, once top of this queue at 4 inbound links, was curated in #9954. Pituitary gland adenoma left the list not by being curated but by being recognised — a skos:narrowMatch on the entry that already covered it (#10069). And several resolved to existing entries the lexical check had missed: velocardiofacial syndrome → 22q11.2_Deletion_Syndrome, proximal spinal muscular atrophy → Spinal_Muscular_Atrophy, RASopathy → RASopathies.

Ranked by inbound links from existing pages. Top of the queue, with the coverage check from the section above already applied:

Links MONDO Disease Status
3 MONDO:0044970 mitochondrial disease scope decision needed — likely a Grouping
2 MONDO:0018564 3p25.3 microdeletion syndrome
2 MONDO:0003709 agoraphobia clear
2 MONDO:0011786 allergic rhinitis clear
2 MONDO:0001164 antisocial personality disorder clear
2 MONDO:0024557 ataxia-telangiectasia-like disorder 1
2 MONDO:0005230 cellulitis clear
2 MONDO:0017843 congenital pulmonary sequestration clear
2 MONDO:0008021 Cowden syndrome NEC risk: eponym + numbered series
2 MONDO:0015474 cryptosporidiosis
2 MONDO:0015612 Dent disease clear
2 MONDO:0001521 intermittent explosive disorder clear

Recommended next tranche — the psychiatric trio. agoraphobia, antisocial personality disorder and intermittent explosive disorder: two inbound links each, all CURATE_ROOT, none in the stub queue, none blocked by a curated ancestor, none with an open issue. Psychiatry is thin in the KB and the three share a DSM/epidemiology literature base, so they cost less together than apart.

Second tranche — the pituitary follow-on. acromegaly (MONDO:0019933) and hypopituitarism (MONDO:0005152), both clear and unclaimed, both sitting on work already in place: Pituitary_Tumor.yaml is now anchored, the somatotroph_camp_pka_overactivation module exists, and Somatotroph_cAMP_PKA_Pituitary_Tumor_Syndromes groups them. Acromegaly should be its own entry that conforms to that module rather than a subtype of pituitary tumour — ectopic GHRH-driven acromegaly involves no pituitary tumour at all.

Do not claim off a stale copy of this list. Regenerate with just gen-dashboard, drop rows curated on current main, then run the coverage audit. Skipping the first step is how the myocarditis miss happened; skipping the third is how #10069 happened.

Workstream 2 — Module-conformance suggestions awaiting triage

66 suggestions at cosine ≥ 0.90; 55 still actionable — 8 have been wired since, and 3 name a node that no longer exists. Data: conformance-suggestions.tsv (full set of 1,128 in research/conforms_to_suggestions.tsv).

The suggestion set itself is stale and cannot currently be regenerated. research/conforms_to_suggestions.tsv dates from 2026-08-12 and has no committed generator, so it predates roughly 600 entries. Only each suggestion's status is re-derived here. Recovering or rewriting the node-embedding generator would be worth more than triaging what it produced a month ago.

39.5% of entries declare a conforms_to (up from 31.2%), and a node-embedding pass proposed links for some of the rest. The high-confidence tail is small enough to review by hand.

These are suggestions, not findings. The embedding matches on node label similarity, so it is blind to direction and to disease context. Two failure modes are already visible in the ≥0.90 set and must be rejected, not applied:

Highest-confidence candidates that do look right:

Each accepted link must satisfy the module's stated conformance bar — the biological processes and causal edges the module expects — not just a matching node name.

Workstream 3 — Lowest-compliance existing entries

Three coherent clusters in the bottom 30. Data: low-compliance-entries.tsv (bottom 60, with the exact missing sections per file). Scores are weighted compliance against 2,535 entries; the KB average is 90.0%.

The epilepsy and cardiac clusters have both improved since the first pass — the worst entry was 53.7% and is now 61.9% — but both persist, and a third has appeared.

Epilepsy and developmental epileptic encephalopathy (12 of the bottom 30)

Juvenile_Myoclonic_Epilepsy (65.8), DNM1_Encephalopathy (67.2), SNAP25_Encephalopathy (69.1), Epilepsy_of_Infancy_with_Migrating_Focal_Seizures (69.3), Pyridoxine-Dependent_Epilepsy (69.8), PCDH19_Clustering_Epilepsy (69.9), PNPO_Deficiency (69.9), Epilepsy_with_Myoclonic_Atonic_Seizures (69.9), Mesial_Temporal_Lobe_Epilepsy_with_Hippocampal_Sclerosis (70.3), KCNQ2_Developmental_and_Epileptic_Encephalopathy (70.3), Febrile_Infection-Related_Epilepsy_Syndrome (70.5), CPLX1-Related_DEE (70.8).

Shared gaps: environmental, histopathology, definitions, epidemiology, experimental_models, classifications, datasets, biochemical. The cluster shares a literature base and the ILAE classification vocabulary, so it is far cheaper as one tranche — classifications in particular is near-mechanical once the ILAE mapping is settled once.

Structural and inflammatory cardiac (6 of the bottom 30)

Tetralogy_of_Fallot (61.9 — now the lowest-scoring entry in the KB), Pericarditis (63.9), Infective_Endocarditis (64.4), Rheumatic_Heart_Disease (66.1), Hypoplastic_Left_Heart_Syndrome (67.9), Familial_Atrial_Fibrillation (69.8).

Shared gaps: datasets, clinical_trials, histopathology, definitions, biochemical. Congenital-heart trials are well registered on ClinicalTrials.gov, so clinical_trials is the cheapest win here.

Molecularly stratified oncology (8 of the bottom 30) — new

Uveal_Melanoma (65.9), FGFR_Altered_Cholangiocarcinoma (69.3), Diffuse_Large_B_Cell_Lymphoma (69.7), Kaposi_Sarcoma (69.9), IDH_Mutant_Cholangiocarcinoma (70.3), IDH_Mutant_Astrocytoma (70.6), RET_Fusion_Thyroid_Cancer (70.8), Li-Fraumeni_Syndrome (70.9).

This cluster did not exist at the first pass and is worth reading together with the over-broad-anchor finding in #10679: the L4 biomarker-stratum entries (IDH_Mutant_AML, and the three colorectal strata) are exactly the entries that anchor a parent MONDO term without the skos:narrowMatch mapping design decisions L4 requires. Low compliance and loose anchoring look like the same underlying thing — these entries were split out of a parent quickly and not finished.

Sturge-Weber_Syndrome (62.7) and Beare-Stevenson_Cutis_Gyrata_Syndrome (65.9) sit in the bottom 10 and belong to no cluster; each needs its own pass.

Workstream 4 — Outstanding entries from the original disease list

177 rows still outstanding of the 431 on initial-diseases.tsv (207 were flagged as candidates; 30 have been curated since that file was written). Data: outstanding-initial-diseases.tsv.

Split by whether a related entry already exists:

Not in scope here (tracked elsewhere)

These are real backlogs but have their own homes:

Regenerating this backlog

just gen-dashboard        # refreshes dashboard/{capability_metrics,not_yet_curated,priority,reports}.json

The four TSVs under projects/CURATION_BACKLOG/ are derived from those dashboards plus research/conforms_to_suggestions.tsv. Note that dashboard/priority.json reports already_curated: 0 / coverage_percent: 0.0 even though the generator log says it excluded 2,932 already-curated roots — the summary appears to count after exclusion, so do not read that 0% as a coverage statement.