VaidhyaMegha/pathways-kg

Pathways Knowledge Graph — partial build

138,725 nodes. 601,322 edges. Human biological pathways, protein interactions, Gene Ontology annotations, genes, drugs and diseases in one graph.

Built with Samyama Graph. Loader: samyama-ai/pathways-kg.

Known gaps

Measured against the engine directly, not inferred:

#	Gap	Detail
1	PARTICIPATES_IN is 2% of its reported size	The Reactome loader reports participates_in_edges=143723. The graph holds 3,423. 140,300 edges the loader says it created do not exist. Confirmed by querying the engine, so this is not an export artefact.
2	29,712 orphan nodes — 21.4% of the graph	Complex 15,711 of 16,169 (97% isolated) · Protein 13,015 of 46,122 (28%) · Pathway 986. Concentrated exactly where gap 1 removes edges.
3	WikiPathways genes unmatched	The WikiPathways phase loaded 986 pathways but reported 42,791 unmatched genes. All 986 of those pathways are therefore isolated — they carry no member edges.
4	Reaction and Compound are zero	Both are defined in the repository schema. Neither was produced by this build. Reactome contributed pathways, proteins and complexes only.

Consequence for the totals: the repository documents ~835K edges; this build has 601,322. Gap 1 accounts for the majority of that 234K difference. The newer source releases used here (see below) add nodes and cannot explain missing edges.

What is sound in this build

• Node and edge totals reconcile exactly against the snapshot header: 138,725 / 601,322.
• 0 dangling edges — every src and tgt resolves to a node present in this dataset.
• 0 duplicate ids within any node file, and 0 ids reused across files.
• INTERACTS_WITH (227,971) and ANNOTATED_WITH (232,365) match their loader-reported counts exactly, as do the GO hierarchy edges. The STRING and Gene Ontology layers appear complete.

So the protein-interaction and GO-annotation halves of the graph are usable; the Reactome pathway-membership and complex-composition halves are not.

Sources and licences

Every licence below was read at the upstream source, not taken from repository documentation.

Source	Version / file	Retrieved	Licence	Verified at
Reactome	current — ReactomePathways.txt, ReactomePathwaysRelation.txt, UniProt2Reactome_All_Levels.txt, reactome.homo_sapiens.interactions.tab-delimited.txt, ComplexParticipantsPubMedIdentifiers_human.txt	2026-08-29	CC BY 4.0	reactome.org/license
STRING	v12.0, organism 9606 — 9606.protein.links, .info, .aliases; confidence threshold 700	2026-08-29	CC BY 4.0	string-db.org/cgi/access
Gene Ontology	current.geneontology.org/ontology/go.json + goa_human.gaf	2026-08-29	CC BY 4.0	GO citation policy
WikiPathways	wikipathways-20260810-gmt-Homo_sapiens.gmt	2026-08-29	CC0	wikipathways.org/terms.html
UniProt	reviewed human proteome (Swiss-Prot), TSV	2026-08-29	CC BY 4.0	rest.uniprot.org/help/license

Required attribution

CC BY 4.0 makes attribution a condition of use for four of the five sources, and it travels to anything you redistribute or build on:

Contains data from Reactome (CC BY 4.0), STRING v12.0 (CC BY 4.0), the Gene Ontology (CC BY 4.0) and UniProt (CC BY 4.0). WikiPathways content is released under CC0. None of these projects endorses this dataset or any analysis derived from it.

STRING additionally asks that you "inform users of any changes or additions that you might have made to the data" — this build applies a confidence threshold of 700 and maps ENSP identifiers to UniProt accessions.

The Gene Ontology asks that a release be cited; this build used the current release as of 2026-08-29 rather than a pinned dated release (see Source substitutions).

⚠️ UniProt rights caveat

UniProt's licence page carries a limitation that must travel with the data:

"We cannot provide unrestricted permission regarding the use of the data, as some data may be covered by patents or other rights."

CC BY 4.0 covers UniProt's own copyrightable content; it does not clear third-party patent or other rights that may attach to particular records. Satisfy yourself of this before commercial use of the UniProt-derived layer (Protein, Gene, Disease, Drug nodes).

Repository licence is separate

The pathways-kg repository code is Apache-2.0. That covers the loader and schema only. It does not govern the data in this dataset, which is bound by the five upstream licences above. license: other is used here because no single SPDX tag expresses "four CC BY 4.0 sources plus one CC0 source".

KEGG is deliberately absent. The repository's design notes record that "KEGG has restrictive licensing (no bulk download, no redistribution). We use open alternatives that collectively exceed KEGG's coverage." That was a licence decision made at design time.

Source substitutions

Two of the five pinned source URLs in the repository are dead, so newer releases were substituted. This is part of why counts differ from the documented figures:

Source	Repository pins	404s?	Used here
WikiPathways	wikipathways-20240310-gmt-Homo_sapiens.gmt	yes	wikipathways-20260810-gmt-Homo_sapiens.gmt
Gene Ontology	release.geneontology.org/2024-06-17/ontology/go.json.gz	yes	current.geneontology.org/ontology/go.json

Both substitutions increase node counts relative to the pinned versions. Neither can account for missing edges.

Contents

nodes/

File	Rows
protein.csv	46,122
goterm.csv	38,092
gene.csv	19,086
complex.csv	16,169
drug.csv	8,486
disease.csv	6,901
pathway.csv	3,869

Pathway = 2,883 Reactome + 986 WikiPathways.

edges/

File	Rows	Connects	Status
annotated_with.csv	232,365	Protein → GOTerm	✅ matches loader report
interacts_with.csv	227,971	Protein → Protein	✅ matches loader report
is_a.csv	57,456	GOTerm → GOTerm	✅
targets.csv	32,793	Drug → Protein	✅
encodes.csv	19,152	Gene → Protein	✅
regulates.csv	8,190	GOTerm → GOTerm	✅
associated_with.csv	7,302	Protein → Disease	✅
part_of.csv	7,058	GOTerm → GOTerm	✅
participates_in.csv	3,423	Protein → Pathway	⚠️ loader reported 143,723
child_of.csv	2,899	Pathway → Pathway	✅
component_of.csv	2,713	Protein → Complex	⚠️ leaves 97% of complexes isolated

Usage

from datasets import load_dataset
proteins = load_dataset("VaidhyaMegha/pathways-kg", "protein")

Given the gaps above, the layers worth using in this build are protein–protein interaction (STRING) and protein–GO annotation:

// Sound: STRING interaction neighbourhood
MATCH (p:Protein {name: $gene})-[:INTERACTS_WITH]-(q:Protein)
RETURN q.name LIMIT 25

// Sound: GO annotation with ontology closure
MATCH (p:Protein)-[:ANNOTATED_WITH]->(t:GOTerm)-[:IS_A*1..3]->(a:GOTerm)
WHERE a.name = $term
RETURN DISTINCT p.name

// NOT SOUND in this build — pathway membership is 2% populated
// MATCH (p:Protein)-[:PARTICIPATES_IN]->(pw:Pathway) ...

Other limitations

• GO annotation coverage. 906,445 annotations were parsed; 232,365 loaded, with 105,433 skipped for having no matching protein. Absence of an annotation here does not mean absence upstream.
• STRING is thresholded at confidence 700 — medium-to-high confidence only. Lower-scored interactions are excluded by design.
• Human only (NCBI taxon 9606).
• Point-in-time build of live sources; see the retrieval dates above.

Not clinical or diagnostic advice.

Citation

Pathways Knowledge Graph — PARTIAL BUILD (138,725 nodes, 601,322 edges).
Built with Samyama Graph. https://huggingface.co/datasets/VaidhyaMegha/pathways-kg
Loader: https://github.com/samyama-ai/pathways-kg
Sources: Reactome (CC BY 4.0); STRING v12.0 (CC BY 4.0); Gene Ontology (CC BY 4.0);
         WikiPathways (CC0); UniProt (CC BY 4.0). All retrieved 2026-08-29.
NOTE: partial build with known gaps — see the dataset card. Not a v1.0 release.