RDF transformation of the PubTator3 BioBrick: entity mentions, extracted relations, and full-text article metadata, keyed to PubMed, MeSH, NCBI Gene, NCBI Taxonomy, Cellosaurus, dbSNP and OMIM.
Per ops/GRAPH_BRICK_STANDARD.md this brick converts every table in its
source. reports/source-coverage.json accounts for all 9:
| Table | Rows | Status |
|---|---|---|
| chemical2pubtator3 | 145,995,476 | converted |
| disease2pubtator3 | 153,863,673 | converted |
| gene2pubtator3 | 70,016,304 | converted |
| species2pubtator3 | 44,237,127 | converted |
| mutation2pubtator3 | 8,605,057 | converted |
| cellline2pubtator3 | 7,830,151 | converted |
| relation2pubtator3 | 35,618,838 | converted |
| BioC | 3,827,600 | converted (article metadata) |
| bioconcepts2pubtator3 | 421,830,050 | excluded — union of the six mention tables |
bioconcepts2pubtator3 is excluded because a full scan of its type column
returns exactly the six entity types that have their own tables, and its
Chemical count (137,001,794) is the stale copy of chemical2pubtator3
(145,995,476). Converting both would double-count every mention.
The previous version of this brick converted relation2pubtator3 only — 1 of
9 tables — which is why the loaded graph held 5% of the chemical–document links
in its source.
<pubmed-article> pubtator:mentions <entity>— the document↔entity link.- Entities resolve to public registries (MeSH, NCBI Gene, NCBI Taxonomy, Cellosaurus, dbSNP, OMIM) so the graph joins the rest of the KG.
- Relations stay reified (
rdf:subject/predicate/object+prov:wasDerivedFrom→ PMID), unchanged from the previous version so existing queries keep working, and additionally contributementionslinks. - Articles carry
dcterms:title,dcterms:date,bibo:doi,bibo:pmid.
go run ./cmd/convert --source /mnt/raid2/biobricks/pubtator/brick --output brick
go test ./...
go run ../../tools/brick-parity check . # table-coverage contractFull build: 459,336,317 rows converted → 752,932,812 distinct triples (~20 min).
NCBI PubTator3 government annotations. The source package declares no explicit license, so artifact publication is blocked pending a reuse determination. Article text is excluded by design — only identifiers, surface forms and PMIDs are emitted.