Europe PMC Scraper (Biomedical Papers, Preprints, Citations)

Search Europe PMC — the superset of PubMed. 43M+ abstracts (MED, i.e. PubMed/MEDLINE), 10M+ open-access full-text articles (PMC), preprints from bioRxiv and medRxiv (PPR), plus patents, Agricola, Chinese biomedical, theses and the NCBI bookshelf. No key, no login, HTTP-only.

Companion to pubmed-articles-scraper (PubMed only), openalex-scholar-scraper (works), crossref-works-scraper (articles) and datacite-scraper (datasets and software).

• search — a Lucene query and/or structured filters (sources, author, journal, publication type, year range, open access, full text in Europe PMC, has-abstract). Cursor-paginated — no result ceiling.
• articles — look up specific articles by PMID (23000000), source-qualified id (MED:23000000, PMC:PMC3531190, PPR:PPR1303040) or DOI → the full core record each.
• citation-network — walk an article's citations forwards (citing — who cites it) or backwards (cited — its reference list).

Set includeAnnotations to attach Europe PMC's text-mined annotation layer — genes/proteins, diseases, chemicals, organisms, cell types, accession numbers, extracted from the full text.

Record type	One per	Carries
SEARCH_SUMMARY	search / citation walk	composed query, hitCount, resultsReturned, pagesFetched, requestsMade, duplicatesSkipped, truncated
ARTICLE	article	id, source, pmid, pmcid, doi, title, abstract, authors (+ ORCID + affiliation), journal, pubYear, citedByCount, isOpenAccess, license, meshTerms, chemicals, keywords, grants, fullTextUrls, textMinedAnnotations, europepmcUrl
ERROR	bad input / missing article	_error + _errorDetail

Every ARTICLE row carries the verbatim Europe PMC record in raw (drop with slimOutput).

Things this API will mislead you about

Each is measured, and each has a scenario in tests/smoke/europepmc-scraper_traps.sh (10/10 passing).

⭐ The citation endpoints page by page and pageSize — and page indexes pages of the size you asked for. Shrinking pageSize on the last page to avoid over-fetching silently re-serves rows you already have: page=1&pageSize=100 then page=2&pageSize=50 returns rows 51–100 a second time, HTTP 200 throughout, with no duplicate signal anywhere in the response. Measured 2026-08-29 — 150 rows of which only 100 were unique. This Actor holds pageSize constant for the whole walk, trims at the end, and reports duplicatesSkipped so any future reshuffle is visible rather than silent.

Every response is HTTP 200 — a malformed query, a gibberish query and a non-existent article all answer 200. There is no "bad query" signal, so a zero-hit search is reported as a SEARCH_SUMMARY with hitCount: 0, never as an error, and an empty run stays distinguishable from a failed one.

An unknown SRC returns zero hits, silently. sources is validated against Europe PMC's controlled list up front rather than becoming a mystery empty result.

/search has no page parameter — deep paging is cursorMark only (*, then nextCursorMark). pageSize maxes at 1000; asking for 1001 returns 200 with an empty result list.

There is no /article/<id> endpoint. A single article is looked up through search?query=EXT_ID:<id> AND SRC:<src> (or DOI:<doi>), which is why a missing article is an empty result list rather than a 404.

A DOI has no citation endpoint. citation-network needs a PMID / MED: / PMC: id; a DOI is rejected before any request instead of returning nothing.

The list fields are {singular: [...]} wrappers that can each be null — meshHeadingList.meshHeading[], grantsList.grant[], chemicalList.chemical[], keywordList.keyword[], authorList.author[], fullTextUrlList.fullTextUrl[], pubTypeList.pubType[]. All are unwrapped to plain lists.

isOpenAccess / inEPMC / hasPDF are the strings "Y" / "N", not booleans. Normalised to real booleans.

id is source-specific — a MED record's id is its PMID, a PMC record's is its PMCID, a PPR record's is PPR…. pmid / pmcid / doi are separate fields, and any of them can be absent.

citations / references rows are lean — title, authors, year, journal abbreviation, id, source, cited-by count. No abstract. Feed the ids back through articles mode if you need the full records.

Notes on cost

search: one request per 100 rows (cursor pages). articles: one request per article, plus one more each if includeAnnotations is on. citation-network: one request per 100 linked articles. resultType: "lite" returns a much smaller row when you only need bibliographic fields. Europe PMC has no anti-bot layer, so no proxy is needed — the default is off.

Policy

www.ebi.ac.uk/robots.txt allows /europepmc/webservices/rest/…. No key, no login, no authenticated surface is touched. Europe PMC content is a mix of licences — each row's isOpenAccess and license say which.