concept string | pmid int32 |
|---|---|
MESH:C000002 | 766,231 |
MESH:C000002 | 5,465,265 |
MESH:C000002 | 5,750,153 |
MESH:C000002 | 5,750,161 |
MESH:C000002 | 5,751,194 |
MESH:C000002 | 5,752,067 |
MESH:C000002 | 5,752,068 |
MESH:C000002 | 6,115,331 |
MESH:C000002 | 34,207,904 |
MESH:C000002 | 37,077,472 |
MESH:C000006 | 396,545 |
MESH:C000006 | 1,312,491 |
MESH:C000006 | 1,334,825 |
MESH:C000006 | 1,425,142 |
MESH:C000006 | 1,478,037 |
MESH:C000006 | 1,478,043 |
MESH:C000006 | 1,484,021 |
MESH:C000006 | 1,533,580 |
MESH:C000006 | 1,547,673 |
MESH:C000006 | 1,592,870 |
MESH:C000006 | 1,679,597 |
MESH:C000006 | 1,680,601 |
MESH:C000006 | 1,688,069 |
MESH:C000006 | 1,779,020 |
MESH:C000006 | 1,814,914 |
MESH:C000006 | 1,819,787 |
MESH:C000006 | 1,827,378 |
MESH:C000006 | 1,828,740 |
MESH:C000006 | 1,855,435 |
MESH:C000006 | 1,932,902 |
MESH:C000006 | 1,936,618 |
MESH:C000006 | 1,967,767 |
MESH:C000006 | 1,979,763 |
MESH:C000006 | 2,017,469 |
MESH:C000006 | 2,142,038 |
MESH:C000006 | 2,142,052 |
MESH:C000006 | 2,188,847 |
MESH:C000006 | 2,192,194 |
MESH:C000006 | 2,199,281 |
MESH:C000006 | 2,202,711 |
MESH:C000006 | 2,247,996 |
MESH:C000006 | 2,418,111 |
MESH:C000006 | 2,500,372 |
MESH:C000006 | 2,510,201 |
MESH:C000006 | 2,553,509 |
MESH:C000006 | 2,566,520 |
MESH:C000006 | 2,656,140 |
MESH:C000006 | 2,667,016 |
MESH:C000006 | 2,668,052 |
MESH:C000006 | 2,668,841 |
MESH:C000006 | 2,677,642 |
MESH:C000006 | 2,680,669 |
MESH:C000006 | 2,685,502 |
MESH:C000006 | 2,691,980 |
MESH:C000006 | 2,693,445 |
MESH:C000006 | 2,852,085 |
MESH:C000006 | 2,866,744 |
MESH:C000006 | 2,866,745 |
MESH:C000006 | 2,870,943 |
MESH:C000006 | 2,906,871 |
MESH:C000006 | 2,951,208 |
MESH:C000006 | 2,951,216 |
MESH:C000006 | 2,964,325 |
MESH:C000006 | 2,975,553 |
MESH:C000006 | 2,975,567 |
MESH:C000006 | 2,975,573 |
MESH:C000006 | 2,998,678 |
MESH:C000006 | 3,042,252 |
MESH:C000006 | 3,044,867 |
MESH:C000006 | 3,046,926 |
MESH:C000006 | 3,052,740 |
MESH:C000006 | 3,053,619 |
MESH:C000006 | 3,053,993 |
MESH:C000006 | 3,062,266 |
MESH:C000006 | 3,065,016 |
MESH:C000006 | 3,066,561 |
MESH:C000006 | 3,066,607 |
MESH:C000006 | 3,066,649 |
MESH:C000006 | 3,069,528 |
MESH:C000006 | 3,071,068 |
MESH:C000006 | 3,071,441 |
MESH:C000006 | 3,073,315 |
MESH:C000006 | 3,083,825 |
MESH:C000006 | 3,085,491 |
MESH:C000006 | 3,086,489 |
MESH:C000006 | 3,094,791 |
MESH:C000006 | 3,105,986 |
MESH:C000006 | 3,109,661 |
MESH:C000006 | 3,139,910 |
MESH:C000006 | 3,161,014 |
MESH:C000006 | 3,281,288 |
MESH:C000006 | 3,282,956 |
MESH:C000006 | 3,284,401 |
MESH:C000006 | 3,298,081 |
MESH:C000006 | 3,299,948 |
MESH:C000006 | 3,300,042 |
MESH:C000006 | 3,304,784 |
MESH:C000006 | 3,313,753 |
MESH:C000006 | 3,315,514 |
MESH:C000006 | 3,319,351 |
PubTator3 entity index for PubMatrix Reloaded
A reformatting of PubTator3's bulk annotations, arranged so that a web browser can read small slices of it with DuckDB WASM. It serves PubMatrix Reloaded, which counts the literature for pairs of biomedical terms.
This is not an NCBI product. The annotations are NCBI's and are unchanged in meaning; only the layout is new. NCBI does not endorse this dataset. For anything beyond this layout, use the original files.
- Source release: 2026-10-06
- Built with:
scripts/build_parquet.py
Files
| Path | Contents |
|---|---|
gene/NNN.parquet |
concept (NCBI Gene ID), pmid |
disease/NNN.parquet |
concept (for example MESH:D001943), pmid |
chemical/NNN.parquet |
concept (for example MESH:D013629), pmid |
names/NNN.parquet |
name, type, concept, n: how often a lowercased name was used for a concept |
relations.parquet |
a, b, type, n: documents with that relation between entities a and b |
shards.json |
For each table, the first value in each shard |
Entity rows are sorted by concept, then pmid, and each (concept, pmid) pair appears once. The names table is built from the bulk files' mentions: names are lowercased in ASCII only, names seen once are dropped, and each name keeps its five most common concepts per type. Concepts sort as bytes. To find a concept, take the last shard whose first concept is not greater than it. In relations.parquet, each pair is stored with a before b, because PubTator3 treats relations as undirected in search.
The data is sharded because DuckDB WASM walks a file's whole row group list for each query, which takes about a second in a large file.
Reading it in a browser
Pass filesystem: { reliableHeadRequests: true, allowFullHTTPReads: false, forceFullHTTPReads: false } to db.open, or DuckDB WASM may download whole files. Query one concept at a time: WHERE concept IN (...) can read far more data. Do not read one file from several workers at once: concurrent range requests for the same URL returned bad bytes in Chromium, seen as ZSTD Decompression failure.
Limits
- Counts differ from the PubTator3 API by a few percent, for reasons not established. Compared on 132 relation cells, the median was 3 to 4% higher here, with rank agreement of at least 0.995.
- Genes are per-species IDs. The API's gene symbols group species together.
- Annotations are machine generated and contain errors. A co-occurrence is a shared document, not evidence of a relationship.
- Only genes, diseases, chemicals, and relations are included: no species, cell lines, or variants, and no free text.
License and citation
NCBI states that this data is a United States Government Work, not subject to copyright, and that NLM and the U.S. Government place no restriction on its use. They disclaim all warranties. See the original notice.
If you use this data, please cite PubTator3:
Wei C-H, Allot A, Lai P-T, Leaman R, Tian S, Luo L, Jin Q, Wang Z, Chen Q, Lu Z. PubTator 3.0: an AI-powered literature resource for unlocking biomedical knowledge. Nucleic Acids Research. 2024;52(W1):W540-W546. doi:10.1093/nar/gkae235
- Downloads last month
- 80