Datasets:

re-cep
/

vitiligo-knowledge-graph

---
license: cc-by-4.0
language:
  - en
pretty_name: Vitiligo Knowledge Graph
tags:
  - biomedical
  - knowledge-graph
  - vitiligo
  - relation-extraction
  - clinical-trials
size_categories:
  - 10K<n<100K
configs:
  - config_name: relations
    data_files: relations.jsonl
    default: true
  - config_name: entities
    data_files: entities.jsonl
  - config_name: evidence
    data_files: evidence.jsonl
  - config_name: ictrp_relations
    data_files: ictrp/relations.jsonl
  - config_name: ictrp_entities
    data_files: ictrp/entities.jsonl
  - config_name: ictrp_evidence
    data_files: ictrp/evidence.jsonl
---
# Vitiligo Knowledge Graph

A knowledge graph of vitiligo research from the
[Vitiligo Initiative](https://github.com/recepsirin/vitiligo-initiative), an
open, non-profit research platform. It links the drugs, genes, diseases,
biomarkers, interventions and clinical trials of the vitiligo literature, and
every relation carries the evidence it rests on: the paper, trial record or
database entry, and where licensing allows, the sentence itself.

| | Count |
|---|---|
| Entities | 17,167 |
| Relations | 93,605 |
| Measured relations | 61,204 |
| Evidence rows | 224,273 |
| Quoted sentences | 85,541 |

A further 2,466 relations resting only on WHO ICTRP records are published
separately under WHO's terms (see [Licensing](#licensing)), bringing the total
to 96,071 relations.

Snapshot of 9 November 2026. Research data, not medical advice.

## Contents

**Entities** (`entities`): 6,899 diseases and adverse events (an adverse
event is the node of the disease or finding it names, so "herpes zoster" in
a trial's safety table and in a paper is one node), 4,348 targets (genes and
proteins), 1,577 trials, 1,386 drugs, 1,201 interventions (procedures,
phototherapy and combination regimens), 912 biomarkers, 807 biological pathways
(WikiPathways CC0, Gene Ontology CC-BY-4.0, and Reactome), 23 outcomes, 10
cutaneous cell types, and related mechanisms and phenotypes.

**Relations** (`relations`): changes 28,592 · associated_with 21,794 ·
participates_in 21,086 · co_expressed_with 8,367 · treats 2,807 · tested_in 2,735 ·
inhibits 2,599 · causes 1,899 · investigates 1,577 · activates 1,139 ·
expressed_in 841 · targets 104 · combined_with 61 · occurs_in 4.

**Evidence** (`evidence`): one row per source of a relation. Incorporates
full-text extraction across 2,712 PMC papers with BioRED BiomedBERT,
direct human vitiligo skin punch biopsy co-expression measurements (GSE311770, N=105),
human skin single-cell RNA-seq lineage expression measurements (GSE298871, 17,057 single cells across 17 biopsies),
WikiPathways (CC0) and experimental Gene Ontology Biological Process (CC-BY-4.0) pathway hierarchies,
and ChEMBL / DrugCentral human drug-target bioactivities ($p\text{Act} \ge 6.5$).
A relation stated by several papers or measurements has several rows.

## Schema

`entities`

| Column | Description |
|---|---|
| `id` | Entity id, referenced by `relations` |
| `kind` | target, drug, disease, trial, biomarker, intervention, ... |
| `key`, `name` | Normalised key and display name |
| `aliases` | Other names |
| `external_ids` | Identifiers in source databases (registry ids, Open Targets, EFO, ...) |

`relations`

| Column | Description |
|---|---|
| `id` | Relation id, referenced by `evidence` |
| `subject_id`, `predicate`, `object_id` | The relation |
| `confidence` | 0-1; registries and curated records rank above text mining |
| `claim` | `asserted` (a source states it) or `observed` (measured; 30,174 relations) |
| `causal_level` | For observed relations, the rung on the causal ladder |
| `extraction_method` | How it was obtained (see [Sources](#sources)) |
| `evidence_count` | Evidence rows for this relation in this release |

`evidence`

| Column | Description |
|---|---|
| `relation_id` | The relation it supports |
| `source` | Where it comes from: `pmid:…`, `ctgov:NCT…`, `opentargets:…`, `reporter:…`, ... |
| `source_kind` | posted_result, relation_extractor, pubtator3, curated_result, expression, trial, prior, gwas, trial_result, mechanism |
| `relation_label` | The text-mining label (BioRED types: Association, Positive_Correlation, ...) |
| `score` | The text miner's score, where there is one |
| `section` | The paper section a sentence came from |
| `text`, `text_licence` | The sentence, where its licence allows; otherwise null |
| `details` | Source-specific fields as JSON (a trial result's numbers, a gene's fold change and FDR, GWAS p-value and cohort, trial phase, ...) |

## Sources

| Method | Relations | How it was made | Measured accuracy |
|---|---|---|---|
| `posted_result` | 27,083 | Results 752 trials posted to ClinicalTrials.gov: adverse events from trials of the same drugs in any disease, outcomes and comparisons from vitiligo trials. Each results group is read as the regimen it received, a drug known by any name MeSH gives it (AIN457 is secukinumab, Stelara ustekinumab); a group switched from placebo to drug, set side by side with another, or pooled is left out | Numbers as posted: 300 of 300 sampled rows match the registry. Groups: 59 of 63 sampled read right |
| `relation_extractor` | 14,776 | Full-text bodies and other text read by the [Vitiligo Relation Extractor](https://huggingface.co/re-cep/vitiligo-relation-extractor), with a loading rule (probability ≥ 0.8, no tables or figures, no co-administration, "treats" only for vitiligo) | 45 of 60 correct (75%); 15 of 17 where vitiligo is one end |
| `pubtator3` | 9,223 | PubTator 3.0's relations from PubMed abstracts, checked against MeSH | Of hand-checked samples: 30 of 34 "treats vitiligo", 31 of 34 "gene associated with vitiligo" |
| `structured` | 5,640 | Trial registries (ClinicalTrials.gov, EU CTR, ISRCTN), Open Targets, and curated trial arms | Database statements |
| `curated_result` | 1,488 | Results read by hand from the papers of 839 trials: responder rates, scores, p-values and adverse events, each with its arm and the section it was read from | Curated; every value cites its place in the paper |
| `expression` | 1,247 | Genes lesional skin expresses differently in three GEO cohorts (GSE65127, GSE75819, GSE311770; a fourth, GSE53146, tested and not used), against the same patients' unaffected skin or healthy donors: limma's moderated t, pairs as pairs, FDR < 5% and a twofold change. A comparison is used only if it sees the melanocytes' pigment genes lower. 1,484 genes in all, 1,247 not otherwise linked to vitiligo | Measured; 73 genes agree across two or more cohorts and rank above those seen once |
| `gwas` | 50 | GWAS Catalog associations, with p-value and cohort | Measured |
| `trial_result` | 21 | Posted trial responder rates | Measured |
| `mechanism_record` | 12 | Curated mechanism hypotheses with their causal rung | Curated |

Trials are the 2,262 vitiligo trials of the corpus: 1,577 here, and the trials
registered only with WHO ICTRP in the `ictrp_*` configurations. Trials known
only from their papers take their interventions from curated trial arms.
Placebo and vehicle arms are recorded as tested, never as treatments.

A `changes` relation joins a regimen to what was measured in the people
given it: what an arm received (a combination such as "NB-UVB + tacrolimus"
is a regimen of its own) and either an outcome family (repigmentation,
relapse, quality of life, ...) or an adverse event. Each evidence row is one
value, with its numbers in `details`, and beside them what the trial's
control group showed: an adverse event row reads "5 of 100 on the drug"
next to "1 of 50 on vehicle". Values pooled across arms name no regimen and
are left out; a placebo or untreated arm appears only as what a regimen was
compared with. Outcomes are taken only from vitiligo trials; adverse events
from trials of the same drugs in any disease.

## Licensing

- **This release (CC BY 4.0).** Relations and pointers are released under
  CC BY 4.0. Quoted sentences are included only where their source permits
  it: 23,191 from open-access papers licensed CC BY or CC0, 588 from NIH
  RePORTER and 214 from ClinicalTrials.gov, whose posted results are
  likewise public records of the US government. Every other sentence is given as
  a pointer to its paper and section; its text is the publisher's or
  author's under the paper's own licence.
- **`ictrp_*` (WHO ICTRP terms).** Relations and evidence resting on records
  of WHO's International Clinical Trials Registry Platform, which permits
  reuse with attribution but not for commercial purposes. Credit WHO ICTRP
  as the source when using them.
- **Excluded.** Eight evidence entries derived from DrugBank, whose terms do
  not allow redistribution.
- Article licences were taken from Europe PMC. BioRED and PubTator 3.0 are
  NCBI resources; Open Targets is CC0; GWAS Catalog data are mostly CC0.
  Expression results are this project's analysis of GEO series, which NCBI
  makes available without restriction.
## Limitations
- Half the relations are **measurements** (30,174): results trials posted
  or printed, and genes expressed differently in lesional skin. The rest are
  **assertions**: a source states them.
- A gene expressed differently in lesional skin is an **association**, not a
  cause: most of the strongest are the genes of the melanocytes the lesion
  has lost.
- A `changes` relation says a regimen was measured against an outcome, not
  that it worked, and an adverse event under a drug is what was recorded in
  the people taking it, not proof the drug caused it: read the numbers
  beside the control group's in its evidence.
- About a quarter of the text-mined relations from full texts are wrong.
  Check a relation against its evidence before relying on it.
- `associated_with` is the most common and least specific relation: it says
  two things are linked, not how.
- A relation's absence means no source in the corpus stated it, not that it
  is false.
## Usage
```python
from datasets import load_dataset
relations = load_dataset("re-cep/vitiligo-knowledge-graph", "relations", split="train")
entities = load_dataset("re-cep/vitiligo-knowledge-graph", "entities", split="train")
evidence = load_dataset("re-cep/vitiligo-knowledge-graph", "evidence", split="train")
```
The release is rebuilt with `vitiligo graph release` from the
[project repository](https://github.com/recepsirin/vitiligo-initiative).
## Citation
If you use this dataset, cite the Vitiligo Initiative and the sources it
builds on: PubTator 3.0 (Wei et al., *Nucleic Acids Research* 2024), BioRED
(Luo et al., *Briefings in Bioinformatics* 2022), Open Targets, the GWAS
Catalog, WHO ICTRP, ClinicalTrials.gov, and the GEO series GSE65127,
GSE75819 and GSE311770.

	---
	license: cc-by-4.0
	language:
	- en
	pretty_name: Vitiligo Knowledge Graph
	tags:
	- biomedical
	- knowledge-graph
	- vitiligo
	- relation-extraction
	- clinical-trials
	size_categories:
	- 10K<n<100K
	configs:
	- config_name: relations
	data_files: relations.jsonl
	default: true
	- config_name: entities
	data_files: entities.jsonl
	- config_name: evidence
	data_files: evidence.jsonl
	- config_name: ictrp_relations
	data_files: ictrp/relations.jsonl
	- config_name: ictrp_entities
	data_files: ictrp/entities.jsonl
	- config_name: ictrp_evidence
	data_files: ictrp/evidence.jsonl
	---

	# Vitiligo Knowledge Graph

	A knowledge graph of vitiligo research from the
	[Vitiligo Initiative](https://github.com/recepsirin/vitiligo-initiative), an
	open, non-profit research platform. It links the drugs, genes, diseases,
	biomarkers, interventions and clinical trials of the vitiligo literature, and
	every relation carries the evidence it rests on: the paper, trial record or
	database entry, and where licensing allows, the sentence itself.

	\| \| Count \|
	\|---\|---\|
	\| Entities \| 17,167 \|
	\| Relations \| 93,605 \|
	\| Measured relations \| 61,204 \|
	\| Evidence rows \| 224,273 \|
	\| Quoted sentences \| 85,541 \|

	A further 2,466 relations resting only on WHO ICTRP records are published
	separately under WHO's terms (see [Licensing](#licensing)), bringing the total
	to 96,071 relations.

	Snapshot of 9 November 2026. Research data, not medical advice.

	## Contents

	Entities (`entities`): 6,899 diseases and adverse events (an adverse
	event is the node of the disease or finding it names, so "herpes zoster" in
	a trial's safety table and in a paper is one node), 4,348 targets (genes and
	proteins), 1,577 trials, 1,386 drugs, 1,201 interventions (procedures,
	phototherapy and combination regimens), 912 biomarkers, 807 biological pathways
	(WikiPathways CC0, Gene Ontology CC-BY-4.0, and Reactome), 23 outcomes, 10
	cutaneous cell types, and related mechanisms and phenotypes.

	Relations (`relations`): changes 28,592 · associated_with 21,794 ·
	participates_in 21,086 · co_expressed_with 8,367 · treats 2,807 · tested_in 2,735 ·
	inhibits 2,599 · causes 1,899 · investigates 1,577 · activates 1,139 ·
	expressed_in 841 · targets 104 · combined_with 61 · occurs_in 4.

	Evidence (`evidence`): one row per source of a relation. Incorporates
	full-text extraction across 2,712 PMC papers with BioRED BiomedBERT,
	direct human vitiligo skin punch biopsy co-expression measurements (GSE311770, N=105),
	human skin single-cell RNA-seq lineage expression measurements (GSE298871, 17,057 single cells across 17 biopsies),
	WikiPathways (CC0) and experimental Gene Ontology Biological Process (CC-BY-4.0) pathway hierarchies,
	and ChEMBL / DrugCentral human drug-target bioactivities ($p\text{Act} \ge 6.5$).
	A relation stated by several papers or measurements has several rows.

	## Schema

	`entities`

	\| Column \| Description \|
	\|---\|---\|
	\| `id` \| Entity id, referenced by `relations` \|
	\| `kind` \| target, drug, disease, trial, biomarker, intervention, ... \|
	\| `key`, `name` \| Normalised key and display name \|
	\| `aliases` \| Other names \|
	\| `external_ids` \| Identifiers in source databases (registry ids, Open Targets, EFO, ...) \|

	`relations`

	\| Column \| Description \|
	\|---\|---\|
	\| `id` \| Relation id, referenced by `evidence` \|
	\| `subject_id`, `predicate`, `object_id` \| The relation \|
	\| `confidence` \| 0-1; registries and curated records rank above text mining \|
	\| `claim` \| `asserted` (a source states it) or `observed` (measured; 30,174 relations) \|
	\| `causal_level` \| For observed relations, the rung on the causal ladder \|
	\| `extraction_method` \| How it was obtained (see [Sources](#sources)) \|
	\| `evidence_count` \| Evidence rows for this relation in this release \|

	`evidence`

	\| Column \| Description \|
	\|---\|---\|
	\| `relation_id` \| The relation it supports \|
	\| `source` \| Where it comes from: `pmid:…`, `ctgov:NCT…`, `opentargets:…`, `reporter:…`, ... \|
	\| `source_kind` \| posted_result, relation_extractor, pubtator3, curated_result, expression, trial, prior, gwas, trial_result, mechanism \|
	\| `relation_label` \| The text-mining label (BioRED types: Association, Positive_Correlation, ...) \|
	\| `score` \| The text miner's score, where there is one \|
	\| `section` \| The paper section a sentence came from \|
	\| `text`, `text_licence` \| The sentence, where its licence allows; otherwise null \|
	\| `details` \| Source-specific fields as JSON (a trial result's numbers, a gene's fold change and FDR, GWAS p-value and cohort, trial phase, ...) \|

	## Sources

	\| Method \| Relations \| How it was made \| Measured accuracy \|
	\|---\|---\|---\|---\|
	\| `posted_result` \| 27,083 \| Results 752 trials posted to ClinicalTrials.gov: adverse events from trials of the same drugs in any disease, outcomes and comparisons from vitiligo trials. Each results group is read as the regimen it received, a drug known by any name MeSH gives it (AIN457 is secukinumab, Stelara ustekinumab); a group switched from placebo to drug, set side by side with another, or pooled is left out \| Numbers as posted: 300 of 300 sampled rows match the registry. Groups: 59 of 63 sampled read right \|
	\| `relation_extractor` \| 14,776 \| Full-text bodies and other text read by the [Vitiligo Relation Extractor](https://huggingface.co/re-cep/vitiligo-relation-extractor), with a loading rule (probability ≥ 0.8, no tables or figures, no co-administration, "treats" only for vitiligo) \| 45 of 60 correct (75%); 15 of 17 where vitiligo is one end \|
	\| `pubtator3` \| 9,223 \| PubTator 3.0's relations from PubMed abstracts, checked against MeSH \| Of hand-checked samples: 30 of 34 "treats vitiligo", 31 of 34 "gene associated with vitiligo" \|
	\| `structured` \| 5,640 \| Trial registries (ClinicalTrials.gov, EU CTR, ISRCTN), Open Targets, and curated trial arms \| Database statements \|
	\| `curated_result` \| 1,488 \| Results read by hand from the papers of 839 trials: responder rates, scores, p-values and adverse events, each with its arm and the section it was read from \| Curated; every value cites its place in the paper \|
	\| `expression` \| 1,247 \| Genes lesional skin expresses differently in three GEO cohorts (GSE65127, GSE75819, GSE311770; a fourth, GSE53146, tested and not used), against the same patients' unaffected skin or healthy donors: limma's moderated t, pairs as pairs, FDR < 5% and a twofold change. A comparison is used only if it sees the melanocytes' pigment genes lower. 1,484 genes in all, 1,247 not otherwise linked to vitiligo \| Measured; 73 genes agree across two or more cohorts and rank above those seen once \|
	\| `gwas` \| 50 \| GWAS Catalog associations, with p-value and cohort \| Measured \|
	\| `trial_result` \| 21 \| Posted trial responder rates \| Measured \|
	\| `mechanism_record` \| 12 \| Curated mechanism hypotheses with their causal rung \| Curated \|

	Trials are the 2,262 vitiligo trials of the corpus: 1,577 here, and the trials
	registered only with WHO ICTRP in the `ictrp_*` configurations. Trials known
	only from their papers take their interventions from curated trial arms.
	Placebo and vehicle arms are recorded as tested, never as treatments.

	A `changes` relation joins a regimen to what was measured in the people
	given it: what an arm received (a combination such as "NB-UVB + tacrolimus"
	is a regimen of its own) and either an outcome family (repigmentation,
	relapse, quality of life, ...) or an adverse event. Each evidence row is one
	value, with its numbers in `details`, and beside them what the trial's
	control group showed: an adverse event row reads "5 of 100 on the drug"
	next to "1 of 50 on vehicle". Values pooled across arms name no regimen and
	are left out; a placebo or untreated arm appears only as what a regimen was
	compared with. Outcomes are taken only from vitiligo trials; adverse events
	from trials of the same drugs in any disease.

	## Licensing

	- This release (CC BY 4.0). Relations and pointers are released under
	CC BY 4.0. Quoted sentences are included only where their source permits
	it: 23,191 from open-access papers licensed CC BY or CC0, 588 from NIH
	RePORTER and 214 from ClinicalTrials.gov, whose posted results are
	likewise public records of the US government. Every other sentence is given as
	a pointer to its paper and section; its text is the publisher's or
	author's under the paper's own licence.
	- *`ictrp_` (WHO ICTRP terms).** Relations and evidence resting on records
	of WHO's International Clinical Trials Registry Platform, which permits
	reuse with attribution but not for commercial purposes. Credit WHO ICTRP
	as the source when using them.
	- Excluded. Eight evidence entries derived from DrugBank, whose terms do
	not allow redistribution.
	- Article licences were taken from Europe PMC. BioRED and PubTator 3.0 are
	NCBI resources; Open Targets is CC0; GWAS Catalog data are mostly CC0.
	Expression results are this project's analysis of GEO series, which NCBI
	makes available without restriction.

	## Limitations

	- Half the relations are measurements (30,174): results trials posted
	or printed, and genes expressed differently in lesional skin. The rest are
	assertions: a source states them.
	- A gene expressed differently in lesional skin is an association, not a
	cause: most of the strongest are the genes of the melanocytes the lesion
	has lost.
	- A `changes` relation says a regimen was measured against an outcome, not
	that it worked, and an adverse event under a drug is what was recorded in
	the people taking it, not proof the drug caused it: read the numbers
	beside the control group's in its evidence.
	- About a quarter of the text-mined relations from full texts are wrong.
	Check a relation against its evidence before relying on it.
	- `associated_with` is the most common and least specific relation: it says
	two things are linked, not how.
	- A relation's absence means no source in the corpus stated it, not that it
	is false.

	## Usage

	```python
	from datasets import load_dataset

	relations = load_dataset("re-cep/vitiligo-knowledge-graph", "relations", split="train")
	entities = load_dataset("re-cep/vitiligo-knowledge-graph", "entities", split="train")
	evidence = load_dataset("re-cep/vitiligo-knowledge-graph", "evidence", split="train")
	```

	The release is rebuilt with `vitiligo graph release` from the
	[project repository](https://github.com/recepsirin/vitiligo-initiative).

	## Citation

	If you use this dataset, cite the Vitiligo Initiative and the sources it
	builds on: PubTator 3.0 (Wei et al., Nucleic Acids Research 2024), BioRED
	(Luo et al., Briefings in Bioinformatics 2022), Open Targets, the GWAS
	Catalog, WHO ICTRP, ClinicalTrials.gov, and the GEO series GSE65127,
	GSE75819 and GSE311770.