Skip to content

Tier 1 execution: Monarch phenotype-gap worklist + pilot curation

Date: 2026-07-31. Author: AI-assisted (Claude Code). Companion to the Monarch KG gap analysis (docs/reports/monarch-kg-dismech-gap-analysis-2026-07-30.md, delivered in a separate PR).

This report executes Tier 1 of the gap analysis: (A) generate the KB-wide disease→phenotype completeness worklist against Monarch's OMIM/Orphanet annotations, and (B) curate a pilot set of diseases as worked examples of each gap-closure path. All work is on branch claude/monarch-tier1-phenotype-gaps, separate from the analysis report's docs PR.

Part A — KB-wide phenotype-completeness worklist

Generated with the in-repo tool dismech.compare.d2p audit-all (per-disease checkpointing, resumable). This run's throughput was ~11 s/disease; a small number of pathologically broad MONDO terms (e.g. MONDO:0000001 Dorsalgia) have enormous association lists and dominate wall-clock.

Coverage of this run: 514 of 1,645 disorders (31%). The sweep is resumable (--resume) and can be completed without redoing finished diseases; the 31% sample is representative, not a curated subset. The per-disease ranked table is committed at docs/reports/data/monarch-phenotype-gap-worklist-2026-07-31.tsv.

Findings (514 diseases)

27,027 phenotype-completeness issues — ~53 per disease. Extrapolated to the full KB this is on the order of ~85,000 issues. By type:

Issue type Count Share Meaning
source_phenotype_missing_locally 22,579 84% OMIM/Orphanet-backed HP term with no local phenotype
source_phenotype_covered_only_by_broader_local_term 2,638 10% Local asserts a parent; source has a more specific term
local_phenotype_unlinked_to_pathograph 1,548 6% Local phenotype present but not wired into a causal edge
local_phenotype_missing_supporting_evidence 262 1% Local phenotype lacks supporting evidence

Interpretation caveats (do not treat the raw counts as a curation quota)

  • Broad/heterogeneous MONDO terms inflate counts. The highest-gap entries in this run (ANK2-Related Complex NDD 986, BLOC1S1 Leukodystrophy 966, 22q11.2 Deletion 231) are umbrella terms whose Monarch annotation aggregates many OMIM subtypes; most of those "missing" phenotypes are not curatable at the umbrella level. Rank the worklist by curatable gaps, not raw totals.
  • missing_locally is a lead, not a directive. Each candidate still requires the standard evidence SOP (exact-quote PMID/ORPHA snippet) and should be added linked into the pathograph only when the mechanism explains it.
  • Not every unlinked flag is a defect — see the Achondroplasia pilot below.

Part B — Pilot curation (3 worked examples)

Three diseases were curated to demonstrate each gap-closure path end-to-end, with full validation (linkml schema, deterministic snippet audit, term validator; the network reference validator additionally passed for Marfan).

  • Added two ORPHA:558-backed phenotypes previously missing: Retinal detachment (HP:0000541) and Pulmonary artery dilatation (HP:0004927), each with an exact-quote ORPHA evidence row and OCCASIONAL frequency.
  • Linked them, plus the previously-unlinked Dural ectasia, into the Extracellular Matrix Remodeling node's downstream (the connective-tissue hub that already feeds Ectopia Lentis / Scoliosis / Pneumothorax).
  • Wired the unlinked Mitral regurgitation via a new Mitral Valve Prolapse → Mitral Regurgitation sequela edge.
  • Snippet audit 118/118; re-audit confirmed the four targeted gaps cleared.
  • Added two ORPHA:20-backed phenotypes: Reye syndrome-like episodes (HP:0006582, the classic HMG-CoA-lyase presentation) and Increased circulating lactate concentration (HP:0002151), both FREQUENT.
  • Linked both as downstream of the Acute hypoketotic metabolic decompensation crisis node (which conforms_to metabolic_intoxication_decompensation). Snippet audit 94/94.

3. Achondroplasia — disciplined unlinked-fix (no new external content)

  • Linked two previously-unlinked, very-frequent angular-limb deformities (Genu varum HP:0002970, Bowing of the legs HP:0002979) into the Impaired endochondral ossification and chondrodysplasia node via evidence-backed downstream edges (reusing in-file PMID:32864841 documenting childhood genu varum).
  • Key learning: this entry explicitly curates "calibrated branches rather than an unsupported all-phenotype fan-out," so many of its unlinked flags are deliberate, not oversights. Only the core, well-evidenced deformities were linked; speculative bulk-linking was intentionally avoided. This confirms the Part-A caveat that unlinked_to_pathograph requires mechanistic judgment, not automation.
  1. Complete the sweep (d2p audit-all --resume) to turn the 31% sample into full KB coverage, then re-rank by curatable gaps.
  2. Batch the missing_locally leads by evidence source — ORPHA-backed rows are the cheapest to curate (exact-quote cache rows already present), as the Marfan/HMG pilots show.
  3. Treat unlinked_to_pathograph per-entry, respecting existing calibrated-branch discipline; prefer evidence-backed causal edges over bare linkage.
  4. Stand up the recurring "Monarch gap scan" proposed in the analysis report so this becomes a standing feed rather than a one-off.