Skip to content

Monarch KG ⇄ dismech disease–phenotype comparison

Date: 2026-07-31 Issue: #7175 — tripartite gap-exchange (dismech ⇄ Monarch KG ⇄ Mondo) Scope: the 1,615 kb/disorders entries with a MONDO primary anchor and curated phenotypes (phenotypes[].phenotype_term). Source: Monarch v3 API, DiseaseToPhenotypicFeatureAssociation edges. Regenerate:

uv run python scripts/kg_phenotype_gap_audit.py --tsv research/kg_phenotype_gap.tsv   # resumable via local cache

The disease→phenotype counterpart of the gene–disease audit. For each disease it diffs dismech's HP terms against the KG's HPOA phenotype annotations for the same MONDO term (exact HP-id match).

Headline

Metric Value
Diseases compared (MONDO + phenotypes) 1,615
…with ≥1 KG phenotype edge 1,349
…with no KG phenotype edge 266
dismech HP assertions 17,015
KG HP assertions 109,102
Overlap (exact) 10,983
kg_only (raw) 98,119
dismech_only 6,032
Exact-match rate of dismech HP terms 64.5%

Read this differently from the gene audit

The phenotype axis is not symmetric with genes, and the raw kg_only must not be read as "dismech gaps":

  • KG HPOA is ~6.4× denser than dismech (109,102 vs 17,015). HPOA exhaustively lists every phenotype ever annotated to a disease; dismech deliberately curates a selective, mechanistically-relevant subset. So a large kg_only is mostly a breadth difference by design, not missing content.
  • Consequently the actionable signals here are the other two: dismech_only (the dismech → KG direction) and the 266 sole-source diseases (below).
  • Exact-ID matching is a lower bound. HPO is deep; a dismech term that is a parent/child of the KG term reads as a mismatch. True semantic agreement is higher than the 64.5% exact rate. Subsumption-aware matching (via the HP hierarchy) is the main follow-up refinement.

Sanity check (highest exact agreement): the well-curated entries line up cleanly — Fanconi Anemia 101 overlap, Alpha Mannosidosis 64/64, Takayasu Arteritis 52/52, Cystinosis 57/57, Granulomatosis with Polyangiitis 63/65. Where dismech is thorough, it matches HPOA nearly term-for-term.

Exact-ID matching is a lower bound. Re-scoring against the HP is_a hierarchy (scripts/kg_phenotype_subsumption.py, run offline from the same cache — no extra API calls) recovers granularity differences a dismech term shares with a parent/child KG term:

Match class Count % of dismech HP
EXACT 10,983 64.5%
+ MORE_SPECIFIC (dismech finer than KG) 870
+ MORE_GENERAL (dismech coarser than KG) 935
= SEMANTIC overlap 12,788 75.2%
UNMATCHED (truly novel dismech HP) 4,227 24.8%

Exact 64.5% → subsumption-aware 75.2% (+10.7 pts). So a tenth of the apparent "disagreement" is just dismech and HPOA describing the same phenotype at different granularity — real agreement, not a gap. The 4,227 UNMATCHED terms are the refined, higher-confidence dismech → KG contribution set (down from 6,032 raw dismech_only); worklist in research/kg_phenotype_subsumption.tsv (unmatched_terms column). It stays concentrated in the same sole-source diseases — Dravet (46 of 49 novel), Long COVID, Multiple Sclerosis, Celiac, IBD, Murine Typhus, Monkeypox, AL Amyloidosis — confirming §A.

Tiers by KG phenotype count

Tier Diseases Reading
no KG phenotype edge 266 dismech is the sole phenotype source → dismech → KG candidates
broad (n_kg > 60) 468 HPOA-dense; contribute 79,393 (81%) of raw kg_only — breadth, not gaps
clean (1–60) 881 clean kg_only = 18,726 candidate enrichments (still breadth-caveated)

A. dismech as sole phenotype source (266) — strongest dismech → KG signal

Diseases where dismech curates phenotypes but the KG's HPOA has none for that MONDO term. Strikingly, these skew toward common/complex and infectious/acquired conditions — exactly where HPOA (built for rare/Mendelian disease) is thin and dismech adds value:

Long_COVID (23 HP), Multiple_Sclerosis (22), Celiac_Disease (22), Ulcerative_Colitis (14), Crohn_Disease, Murine_Typhus (23), Hantavirus_Pulmonary_Syndrome (16), Monkeypox (14), Neurosarcoidosis (14), Organophosphate_Poisoning (14), AL_Amyloidosis (13), Esophageal_Atresia (15).

These are the cleanest whole-disease contributions to hand to Monarch/HPOA.

B. dismech_only — phenotypes dismech curates that the KG lacks (6,032)

Per-disease phenotype-level contributions (or subsumption artifacts to verify). Top contributors:

  • Dravet_syndrome (MONDO:0100135): 48 dismech-only vs 1 KG phenotype — the MONDO term is essentially unannotated in HPOA; dismech has a full phenotype set.
  • Fanconi_Anemia: 38 dismech-only on top of 101 overlap — rich in both directions.
  • Hypochondroplasia (26), Crohn_Disease (24), COPA_Syndrome (18), Cardiospondylocarpofacial_Syndrome (20), MED13_Syndrome (17).

Some of the 6,032 are granularity mismatches (dismech's term is a parent/child of an HPOA term) rather than true novel annotations — the subsumption-aware refinement above separates these, leaving 4,227 truly-novel terms. Treat that refined set as the higher-confidence dismech_only contribution feed.

C. Clean kg_only — candidate phenotype enrichments (18,726)

For the 881 clean-tier diseases, phenotypes HPOA annotates that dismech doesn't yet carry. Given dismech's selective-curation design this is optional enrichment, not a defect — useful as a curator prompt ("did we omit a mechanistically relevant phenotype?") rather than a gap to close wholesale. Examples with modest, reviewable lists: Blau_Syndrome, Niemann-Pick_Disease_Type_B, neuroferritinopathy, Creatine_Transporter_Deficiency, SETBP1_Disorder (see the TSV's kg_only column).

Caveats

  • Exact-ID match in this table is a lower bound — the subsumption-aware refinement section (and scripts/kg_phenotype_subsumption.py) accounts for parent/child agreement, raising exact 64.5% → semantic 75.2%.
  • HPOA breadth ≠ dismech scope — raw kg_only is a breadth difference by design; never read it as a gap count.
  • Broad anchors still noisy — the 468 n_kg>60 entries include broad/grouping anchors (as in the gene and anchoring audits); fix the anchor before reading their kg_only.

Machine-readable worklist

research/kg_phenotype_gap.tsv — one row per disease: disorder, mondo_id, n_dismech, n_kg, n_overlap, kg_only, dismech_only (HP lists as HP:id(label)).

Follow-ups (#7175)

  • ✅ Subsumption-aware matching (HP hierarchy) — done (kg_phenotype_subsumption.py; see the refinement section above). Exact 64.5% → semantic 75.2%.
  • Feed §A (266 sole-source) and the 4,227 subsumption-UNMATCHED terms into a dismech → Monarch/HPOA contribution set.
  • Companion flows delivered in this PR: groupings/modules anchoring (mondo-anchoring-audit-2026-07-30.md §E) and Mondo → dismech (mondo-to-dismech-gaps-2026-07-31.md).