VaidhyaMegha/clinicaltrials-kg

Clinical Trials Knowledge Graph

7,628,735 nodes. 15,531,427 edges. 575,778 clinical trials registered on ClinicalTrials.gov — with their arms, outcomes, sites, sponsors, conditions, interventions, MeSH codes and linked PubMed publications.

Built with Samyama Graph. Loader and ETL: samyama-ai/clinicaltrials-kg.

Stored as Parquet — 623 MB, against roughly 7 GB as raw JSON.

Attribution — required

Two acknowledgements travel with this data:

Courtesy of the U.S. National Library of Medicine.

Aggregate Analysis of ClinicalTrials.gov (AACT) Database. Clinical Trials Transformation Initiative (CTTI). Available at: https://aact.ctti-clinicaltrials.org/

Neither NLM nor CTTI endorses this dataset, its processing, or anything derived from it.

Licence

Registry data	U.S. Government public domain (ClinicalTrials.gov, via AACT)
PubMed, MeSH	NLM terms — no fees or royalties, commercial use permitted
Loader code	Apache-2.0, in the GitHub repo. This does not govern the data.

license: other is used deliberately rather than a convenient SPDX tag. No single standard licence describes this: the underlying registry content is US Government material outside copyright, while NLM and CTTI both request attribution, which CC0 would wrongly imply is waived. The terms in full:

• NLM (terms) — "No charges, usage fees or royalties"; reuse including commercial is permitted; you must acknowledge NLM, must not suggest NLM endorses your work, and must either track the current version or state that your copy may be outdated. This copy is a snapshot of 2026-04-02 and is not updated — see Limitations.
• AACT / CTTI (site) — AACT is a daily reload of ClinicalTrials.gov: "Content is downloaded from ClinicalTrials.gov daily and loaded into AACT." CTTI states no licence and imposes no restriction, "encourages the use of all materials", and asks that the source be acknowledged, supplying the citation above.

What we checked, and what we did not. Both sets of terms above were read directly at source. Neither prohibits redistribution. CTTI publishes no formal licence grant, so the basis for redistributing the AACT-derived portion is that its content is US Government registry data outside copyright, plus CTTI's own encouragement to use it. That is an assessment from published terms, not legal advice.

⚠️ What was removed, and why

The upstream repository builds this graph from more sources than are published here. Two were excluded because their terms do not permit redistribution:

Removed	Why	Nodes
AdverseEvent	Every record is MedDRA-coded (source_vocabulary: "MedDRA"). MedDRA is licensed by ICH/IFPMA on a paid subscription; its terminology may not be redistributed without one. FDA publishing FAERS does not transfer MedDRA's rights.	145,711
Drug	Every record carries a drugbank_id. DrugBank has tiered terms, some of which prohibit redistribution outright.	5,256

150,967 nodes (1.9%) and 11,448,933 incident edges (42%) were removed. The edge count falls far more than the node count because adverse-event reporting is dense — each trial reports many events. The REPORTED and CODED_AS_DRUG edge types are therefore absent entirely.

No .sgsnap snapshot is shipped with this dataset, unlike the others in this organisation. The snapshot still contains the MedDRA and DrugBank records, so publishing it would undo the exclusion.

If you need adverse-event or drug data, obtain it from MedDRA and DrugBank directly under their own terms, and rebuild with the ETL in the repository.

Files

Nodes carry an id; edges reference those ids as src and tgt.

nodes/

File	Rows	Columns
outcome.parquet	3,564,582	id, description, measure, time_frame, trial_nct_id, type
armgroup.parquet	1,055,533	id, label, trial_nct_id, type
site.parquet	1,029,229	id, city, country, facility
publication.parquet	750,114	id, pmid, reference_type, title
clinicaltrial.parquet	575,778	id, nct_id, title, official_title, brief_summary, phase, overall_status, study_type, enrollment, start_date, completion_date, primary_completion_date, has_results, why_stopped, last_updated
intervention.parquet	472,275	id, description, name, type
condition.parquet	125,689	id, name
sponsor.parquet	49,519	id, class, name
meshdescriptor.parquet	6,016	id, name

edges/

File	Rows	Connects
measures.parquet	3,564,582	ClinicalTrial → Outcome
conducted_at.parquet	3,409,901	ClinicalTrial → Site
coded_as_mesh.parquet	2,623,299	Condition → MeSHDescriptor
uses.parquet	1,282,300	ClinicalTrial → Intervention
has_arm.parquet	1,055,533	ClinicalTrial → ArmGroup
studies.parquet	1,025,129	ClinicalTrial → Condition
published_in.parquet	1,021,552	ClinicalTrial → Publication
tests.parquet	973,353	ArmGroup → Intervention
sponsored_by.parquet	575,778	ClinicalTrial → Sponsor

Usage

from datasets import load_dataset

trials = load_dataset("VaidhyaMegha/clinicaltrials-kg", "clinicaltrial", revision="v1.0")
print(trials["train"].num_rows)      # 575778

Parquet reads well with anything columnar:

import pyarrow.parquet as pq
t = pq.read_table("nodes/clinicaltrial.parquet", columns=["nct_id", "phase", "overall_status"])

// Sponsors running the most trials
MATCH (t:ClinicalTrial)-[:SPONSORED_BY]->(s:Sponsor)
RETURN s.name, count(t) AS trials ORDER BY trials DESC LIMIT 5

// Conditions studied across the most countries — three hops, and the reason
// this is a graph rather than a set of tables
MATCH (c:Condition)<-[:STUDIES]-(t:ClinicalTrial)-[:CONDUCTED_AT]->(site:Site)
RETURN c.name, count(DISTINCT site.country) AS countries, count(DISTINCT t) AS trials
ORDER BY countries DESC LIMIT 10

Verification

• Node and edge totals reconcile against the snapshot header: 150,967 dropped + 7,628,735 written = 7,779,702; 11,448,933 dropped + 15,531,427 written = 26,980,360.
• 0 dangling edges — every src and tgt resolves to a node in this dataset.
• 0 orphan nodes — despite removing 42% of edges, no node was left isolated.
• Distributions match the public registry: 439,465 interventional vs 134,313 observational (~76% interventional), and the highest-volume sponsors come out as Assiut University, Cairo University, GlaxoSmithKline, the National Cancer Institute and AP-HP — which is what ClinicalTrials.gov itself shows.

Limitations

• A 2026-04-02 snapshot, not a live mirror. ClinicalTrials.gov changes daily; NLM's terms require this to be stated. For current data go to the registry or AACT.
• No adverse events and no drug normalisation — see What was removed above. Any analysis of trial safety outcomes is out of scope for this dataset.
• Text fields are as-registered. Titles, summaries and outcome measures are sponsor-written free text, with the inconsistency that implies.
• Sites are facility strings, not geocoded or reconciled to any institution registry.
• Conditions and interventions are strings too; CODED_AS_MESH gives partial normalisation for conditions only.
• Registry data reflects what sponsors registered, which is not always what was conducted or published.

Not medical advice, and not a substitute for the official registry record.

Citation

Clinical Trials Knowledge Graph, v1.0 (7,628,735 nodes, 15,531,427 edges).
Built with Samyama Graph. https://huggingface.co/datasets/VaidhyaMegha/clinicaltrials-kg
ETL: https://github.com/samyama-ai/clinicaltrials-kg
Source data: ClinicalTrials.gov via the AACT database, Clinical Trials Transformation
             Initiative (CTTI), https://aact.ctti-clinicaltrials.org/
             Courtesy of the U.S. National Library of Medicine.
Snapshot date: 2026-04-02

The three pieces

Piece	Where
Code — ETL, schema, loaders	github.com/samyama-ai/clinicaltrials-kg
Data — this dataset, and the registry it derives from	here, and aact.ctti-clinicaltrials.org
Graph — rebuild from these Parquet files with the ETL	(no snapshot shipped — see above)