What querying rare-disease names at scale showed

Why this exists

Rare disease reference data is built for looking up one condition at a time. Querying thousands of official names against literature and trial registries surfaces mismatches that are invisible at single-record scale. These notes are findings about that scale problem. They depend on Orphanet as an open public nomenclature — the comparisons are possible because the data can be read systematically.

What “11,645” and “about 7,000” each mean

Orphanet product1 in this freeze contains 11,645 nomenclature rows, not 7,515 Disorder-level clinical entities. Rows classify as:

• 7,515 Disorder
• 1,457 Subtype of disorder
• 2,673 Group of disorders

The commonly cited figure of roughly 7,000 rare diseases aligns with the Disorder level alone (7,515 in this dump), not with groups or every subtype. This atlas drops groups, then drops preferred names marked OBSOLETE: or NON RARE IN EUROPE: (801 such names among non-group rows), leaving about 8,171 usable entities. The sample of 300 is drawn from that usable set.

1. Official nomenclature and publication/registry language diverge

In the sample of 300, 43 preferred names (14.3%) returned no results in either Europe PMC or ClinicalTrials.gov under phrase and MeSH queries (when MeSH resolved). ClinicalTrials.gov also tried recall-expansion terms; those still returned nothing for this set. That is not evidence that research is absent; it is evidence that those official strings do not function as search terms in those databases.

Separately, publication metrics use a stricter credibility set: 183 of 300 (61.0%) remain in the publications denominator. The other 117 (39%) are excluded because the name does not reliably retrieve literature under our rules (including name-collision / neglect flags and queries that return nothing). About 39% of sampled entries therefore do not support a clean literature count from the preferred label alone.

Orphanet's preferred-label curation is excellent by ordinary catalogue standards: among 11,645 product1 rows we verified only 3 clear character-level misspellings in preferred names (0% of rows). The searchability gap is therefore not explained by typographical quality. Official names and the names used in papers and trial registrations behave like different vocabularies.

Examples of names that returned nothing in either database in this sample:

• ORPHA:1129 Arachnodactyly-abnormal ossification-intellectual disability syndrome
• ORPHA:1174 Cerebellar ataxia-ectodermal dysplasia syndrome
• ORPHA:1564 Dandy-Walker malformation-facial hemangioma syndrome
• ORPHA:2269 Ichthyosis-alopecia-eclabion-ectropion-intellectual disability syndrome
• ORPHA:2408 Lowe-Kohn-Cohen syndrome
• ORPHA:2519 Microcephaly-seizures-intellectual disability-heart disease syndrome
• ORPHA:3194 Corneodermatoosseous syndrome
• ORPHA:3224 Deafness-genital anomalies-metacarpal and metatarsal synostosis syndrome

Reproduce: load findings-2026-07-26.json, field searchability; per-disease query strings are on each disease page.

2. 801 corpus entries are housekeeping labels

Before sampling, this atlas excludes preferred names that begin with OBSOLETE: or NON RARE IN EUROPE: — 801 such rows among non-group entities in this product1 freeze. Those prefixes are operational markers in the nomenclature file. A literal search for a string such as "NON RARE IN EUROPE: Tourette syndrome" returns nothing in the literature databases, which would incorrectly describe a well-studied condition as unresearched if the prefix were queried as part of the name.

3. Hyper-specific labels behave like index entries, not search names

Long multi-feature syndrome strings and <GENE>-related … constructions often return no (or almost no) publications under the preferred label, while a broader Mondo parent concept still has substantial indexed literature. In this sample, 17 entities had zero publications under the preferred name yet an informative parent probe with at least 50 hits. Examples:

• ORPHA:2269 Ichthyosis-alopecia-eclabion-ectropion-intellectual disability syndrome
  Preferred-label publications: 0; Mondo parent “ectodermal dysplasia syndrome”: 690 publications in the parent probe.
• ORPHA:1174 Cerebellar ataxia-ectodermal dysplasia syndrome
  Preferred-label publications: 0; Mondo parent “ectodermal dysplasia syndrome”: 690 publications in the parent probe.
• ORPHA:466921 Childhood-onset progressive contractures-limb-girdle weakness-muscle dystrophy syndrome
  Preferred-label publications: 0; Mondo parent “progressive muscular dystrophy”: 2,712 publications in the parent probe.
• ORPHA:308621 Glycogen storage disease due to glycogen branching enzyme deficiency, progressive hepatic form
  Preferred-label publications: 0; Mondo parent “disorder of glycogen metabolism”: 117 publications in the parent probe.

The sample also contains 12 labels matching a GENE-related … pattern, including cases that return nothing under that exact string.

4. Measurement sensitivity: parent category versus specific name

Preliminary. Trial relevance has dual-model adjudication on a 33-disease gold set; full human validation of trial relevance is not yet complete. Treat the sensitivity range as preliminary.

As of 26 July 2026, among 255 diseases with usable interventional-trial matching:

• 151 of 255 (59.2%) had no interventional trial under the specific condition name
• 125 of 255 (49%) still had none if broader parent-category registrations were treated as filling a zero

As of 26 July 2026, 49% of the 255 diseases sampled had no registered interventional trial when parent-category hits counted; 59.2% had none under the specific condition name alone.

The 49%–59.2% spread is a finding about measurement difficulty: whether a subtype “inherits” its parent disease’s trials is a judgment, and both answers change the headline in the same underlying data. Disease pages show both tiers when a parent category exists.

5. Method fragility: a single character can erase retrievability

Character-level misspellings in preferred labels are rare — 3 verified upstream cases among 11,645 product1 rows (0%). That rate is consistent with careful curation. They matter here as an illustration of method fragility under the nomenclature finding: exact string matching has no tolerance for a missing letter.

ORPHA:436159 shows this clearly. The preferred label reads “Autoimmune lymphoproliferative syndrome due to CTLA4 haploinsuffiency” (one character short of “haploinsufficiency”). One synonym repeats the same spelling; another spells it correctly — so retrieval is partly rescued by synonym coverage the preferred label does not control. Verified upstream (not introduced by our parser):

• ORPHA:693869 Gallblader arteriovenous malformation — synonyms include Arteriovenous malformation of gallbladder
• ORPHA:53583 Paroxysmal dystonic choreathetosis with episodic ataxia and spasticity — synonyms include DYT9; Episodic choreoathetosis/spasticity
• ORPHA:436159 Autoimmune lymphoproliferative syndrome due to CTLA4 haploinsuffiency — synonyms include ALPS due to CTLA4 haploinsuffiency; CHAI; CTLA-4 haploinsufficiency with autoimmune infiltration disease

Methods

Sources: Orphanet / Orphadata product1 and prevalence products; Mondo hierarchy and exact synonyms; GenCC gene–disease validity; Europe PMC for publications; ClinicalTrials.gov API v2 for studies. Versions are those recorded in the snapshot header above.

Sampling: uniform random draw of 300 entities with seed 42 from usable non-group Orphanet rows after dropping OBSOLETE: / NON RARE IN EUROPE: preferred names.

Queries: quoted preferred-name and synonym phrases; MeSH descriptor labels when Mondo xrefs resolve; gene symbols and carefully filtered recall expansions for trials. Names that return nothing on both literature and trials are excluded from site-wide denominators.

Two-tier trials: the headline numerator counts only interventional studies matched to the specific condition name. A separate parent-category tier (singular Mondo name-parent, when present) is shown on disease pages and used only for the sensitivity comparison above. See also How we count trials.

Limitations

• Findings are from a 300-disease sample, not the full usable corpus of ~8,171.
• Trial relevance is not yet fully human-validated; the sensitivity range is preliminary.
• Publication counts are noisy (indexing delays, name variants, full text versus structured fields).
• Matching errors run in both directions: false positives and false negatives.
• The India NPRD policy mapping on disease pages is hand-curated and may be incomplete.

Reproduce this

Numbers on this page come only from the dated freeze findings-2026-07-26.json, built from data/diseases.sample300-seed42.json and data/defect-provenance.json. New measurements write a new dated file; existing snapshots are never overwritten. Defect provenance: npx tsx scripts/verify-defects.ts (reads cached Orphanet XML only). Repository licence: Apache-2.0.

Corrections log

Snapshots are append-only. Corrections ship as a new dated freeze, not by editing this file in place.