v1.0.0·5,504,456 records·0 files·1.6 GiB
Nov 13, 20262
README

System Public Graph

What the published literature reports about relationships between medical and biological concepts, and the studies each report rests on. Five record types, 5,504,456 records.

Record types

Type Records References
Association 1,999,877 studyId, variable1Id, variable2Id, variable1ConceptId, variable2ConceptId
Variable 1,774,551 conceptId
Study 844,435 none
Relationship 839,872 sourceConceptId, targetConceptId
Concept 45,721 none

Every reference resolves inside the collection. They are declared with x-ref-type, so tools can follow them without knowing anything about this dataset in advance.

How the types fit together

An Association is one statistic from one paper: an odds ratio, hazard ratio, mean difference or correlation between two measured variables, with its confidence interval and p-value where the paper reported them.

A Variable is a specific thing a study measured, in the study's own wording ("dysphagic stroke cases", "maximum lifetime body mass index"). A Concept is what such a variable is about ("stroke", "body mass index"). Concepts carry a Wikidata identifier, which is the one field in this collection that points outside it.

A Study is the paper, with its DOI and PubMed identifier.

A Relationship is System's aggregate edge between two concepts, carrying how many findings and studies sit behind it, and the date of the most recent publication on it.

Because the types reference each other, a claim can be read at three levels of detail: concept to concept, the specific variables two papers measured, or the paper itself.

What is in this collection

The complete corpus: all five files, every row.

Source

The Harvard Dataverse release of the System public graph, doi:10.7910/DVN/7V6JDO: five CSV files, 5,504,456 rows, 1.1 GB. Place them in input/. Nothing is downloaded automatically.

License: CC BY-NC 4.0, as System released it. Credit System and the Dataverse release when you use it, and do not use it commercially.

The release is dated 2026-07-08, but its last_updated values run only to December 2024. System also operates a public API beta which may carry newer data. This collection does not track it.

Notes on the data

Things a consumer will run into. None of them are ingest artifacts.

  • Self-edges exist. stroke links to stroke, because two distinct measured variables can map to the same concept.

  • study_type is mostly blank. It reads Experimental Trial on 89,728 of 844,435 studies (10.6%) and is empty otherwise. A blank value does not mean the study was observational, only that the type was not recorded.

  • There is no publication date per study. The only date in the corpus is lastPublishedDate on Relationship, which is the most recent publication on that edge rather than a date for any individual paper. It is populated on roughly 817,000 of 839,872 relationships and spans 1925 to

  • Aggregate counts differ slightly from the associations. Counting the associations on a concept pair does not always match that pair's Relationship.numFindings. For body mass index and mortality it is 1,544 against 1,542. The difference is upstream.

  • Over half the associations report no confidence interval, and coverage depends heavily on the statistic. 52.2% are missing at least one bound, and 31.4% report no p-value.

    Statistic Associations Missing an interval
    hazard_ratio 235,030 16%
    relative_risk_ratio 104,104 20%
    odds_ratio 577,611 21%
    coefficient 281,686 78%
    mean_difference_unstandardized 444,825 94%
    pearson_r 193,449 98%

    This matters for anything that decides significance from the interval. A missing interval means the paper did not report one, or System did not capture it. It does not mean the result was null. Ratio measures can mostly be judged this way; correlations and mean differences almost never can.

  • Two confidence bounds are left out. One association's lower bound of -10,000,000,000,000,000 and another's upper bound of 423,000,000,000,000,000 are integers too large for Underlay to store exactly (beyond ±2^53), and look like placeholders. Both fields are absent on those two records.

Ingest

# The complete corpus, streamed from the CSVs.
npx tsx sources/system/public-graph/ingest.ts push --all --env staging

# A topic subset, useful for testing changes quickly.
npx tsx sources/system/public-graph/ingest.ts plan  --topic cardiometabolic
npx tsx sources/system/public-graph/ingest.ts build --topic cardiometabolic
npx tsx sources/system/public-graph/ingest.ts push  --env staging

push --all counts the corpus first using the same generator it then streams from, so a miscount surfaces in seconds rather than after an hour of uploading. It writes no intermediate file.

A topic subset is closed over its references rather than cut at a row limit. The five types reference each other five ways, so truncating any file partway through would leave associations pointing at variables that were never included.

The full push took 81 minutes for 5,504,456 records at roughly 2,000 records per second.