Checking Whether a Paper Has Been Retracted

What re-checking the dihexa record returned

The Vialog compound record for dihexa states that the paper reporting direct binding to hepatocyte growth factor was retracted in April 2025. Re-checked against PubMed on 17 August 2026, the statement is accurate and incomplete. Widening the query from the single paper named in the secondary literature to the whole cluster of work from the originating laboratory returns four flagged records in one journal, and three retractions rather than one. The record understated the position. The correction itself was narrow; the original check had been run against a single citation rather than against the laboratory's cluster of papers in that journal, and the cluster was where the other three flags sat.

All four papers appeared in the Journal of Pharmacology and Experimental Therapeutics. Kawas and colleagues 2011 (339:509-518, PMID 21859930), Kawas and colleagues 2012 (340:539-548, PMID 22129598), McCoy and colleagues 2013 (344:141-154, PMID 23055539) and Benoist and colleagues 2014 (351:390-402, PMID 25187433) each received a notice printed in September 2021 under the journal's own heading, Notice of Concern, at 378(3):311 through 314, PMIDs 34551987 to 34551990. PubMed indexes all four under the publication type Expression of Concern, and that is the term used below. Three of the four were then retracted in April 2025, the notices appearing at 392(4):103566, 103567 and 103568, PMIDs 40312092 to 40312094.

McCoy 2013 is the exception, and the awkward one. It carries the 2021 expression of concern and nothing since; PubMed lists no retraction publication type against it. That paper describes the synthesis, the metabolic-stability measurements and the behavioural readouts, so the three withdrawn papers supplied the proposed mechanism while the one still standing supplied the molecule and the behaviour. Its flag has been unresolved for close to five years.

Three notice types, three different statements

The three notice types that appear on these records assert different things. Corrections, in the ICMJE formulation, follow honest errors: a correction notice is published detailing the changes and citing the original, and the article remains in the record. Retraction covers two routes in the same formulation: errors serious enough to invalidate a paper's results and conclusions may require it, and a finding of scientific misconduct calls for it outright. An expression of concern asserts neither innocence nor invalidity. ICMJE places it in the section dealing with misconduct, where editors may publish one pending the outcome of an investigation, and it records that a question has been raised and not settled, which is why it can sit on a record for years without resolving in either direction.

Kawas 2011 carries all three. Its PubMed record links an erratum at 351(3):685 in December 2014, the September 2021 expression of concern, and the April 2025 retraction, three notice types on one article across eleven years. A check that read only the newest notice returned the article's history incompletely, and one that read only the oldest returned its status wrong. NISO's recommended practice RP-45-2024, published on 27 June 2024, exists because of exactly this problem, and its stated premise is that these features must be evident to machine-reading and automated processes and not only to human readers.

The flag sits on the record, not on the file

The National Library of Medicine handles this by publication type and by linkage. A retraction notice is typed Retraction Notice and the withdrawn article is typed Retracted Publication; the two citations are linked in both directions, so the article displays a retraction-in line and the notice displays a retraction-of line. Expressions of concern are linked the same way. That linkage is what makes status queryable at all. It attaches to the citation. It does not attach to a PDF in a folder, on a mirror, or in a reference manager's attachment store, and a saved file carries whatever it carried on the day it was saved.

Suelzer and colleagues measured how far the flag actually travels. Their 2021 cross-sectional study took 150 retracted articles, three from each of the fifty journals with the most retracted publications indexed in PubMed, and checked publisher sites and six bibliographic databases. PubMed and the other MEDLINE interface tested met all five criteria 87% of the time (650 of 750); a psychology abstracts index managed 23% (7 of 30) and a nursing and allied health index 7% (8 of 115), the platforms being named in the source. That study itself carries a 2022 correction to the table these figures are drawn from; the corrected cells concern two other databases and the values quoted here are unchanged. On publisher sites, 112 of 150 articles (75%) followed at least six of the ICMJE's seven recommendations, and nine (6.0%) used no label at all. Retraction information appeared consistently in the abstracts for 78% of journals (39 of 50) and in the PDFs for 64% (31 of 50).

Citation does not stop when the notice publishes

Budd, Sievert and Schultz reported in JAMA in 1998 that 235 retracted articles identified in MEDLINE through August 1997 were cited 2,034 times after their retraction notices; of 299 of those citations examined, 19 noted the retraction and 280 treated the work as valid, 263 of them implicitly. Zhu, Jia and Leung's 2024 systematic review in JAMA Network Open analysed 10,461 microRNA biomarker articles, 887 of them retracted. Those were cited 6,327 times after retraction; 621 of the 792 cited retractions (78.41%) picked up at least one post-retraction citation, and 238 (30.05%) were cited more often after retraction than before.

Frasco and colleagues found the same pattern inside a single specialty in 2022: across nine anaesthesiology journals, 245 retracted publications, of which the 211 retracted before January 2020 accumulated 1,307 post-retraction citations in the citation index used by that study, a median of 3.5 each and a maximum of 88. Eighty per cent of those citations affirmed the validity of the retracted work and 5.2% acknowledged the retraction or the misconduct. The rate is not fixed. Suelzer and colleagues 2019 found that among works citing the 1998 Wakefield article, 123 of 322 published between 2005 and 2010 documented the partial retraction (38.2%), rising to 360 of 502 (71.7%) between 2011 and 2018.

What the mechanical check returned on this cluster

The mechanical check ran in three steps, with no judgement in any of them. The PubMed record was retrieved and its publication types and linked comments-and-corrections fields read. Crossref's REST API was queried on the DOI and the update-to and updated-by arrays read. The Retraction Watch database was queried. Run against these four DOIs on 17 August 2026, Crossref returns the three retracted articles with a RETRACTED prefix on the title and McCoy 2013 without it, the same prefix Suelzer's 2021 discussion had proposed as a consistent visual signal. Each of the three also carries an update-to entry pointing at its own DOI, typed as a retraction and dated to the article's original month of publication, so a check that reads that date reads 2011, 2012 or 2014.

Two further oddities surface in the same response. The April 2025 notices appear in each record's updated-by array typed as erratum rather than retraction. The four 2021 expressions of concern appear there sourced not from the publisher but from retraction-watch, carrying record identifiers 32544 to 32547, which means those flags exist in Crossref only because Crossref acquired the Retraction Watch database from the Center for Scientific Integrity on 12 September 2023 and released it as an open file.

That file, retrieved on 17 August 2026, holds 71,844 rows: 66,317 retractions, 3,586 expressions of concern, 1,505 corrections, 160 reinstatements and 276 rows with the nature field left empty, with a most recent retraction date of 8 August 2026. It contains all four September 2021 expressions of concern for these papers, each carrying the reason field Concerns/Issues about Image. It contains no row for the three April 2025 retractions, and of its 21 rows from this journal none is dated 2025. A check run against that database alone therefore returns an expression of concern for three papers PubMed has flagged as retracted for sixteen months.

What the notices themselves yielded

One of the three retraction notices is readable in full; the other two are not. The journal's article page and the publisher's platform each returned HTTP 403 to an unauthenticated request on 17 August 2026, and only one of the three notices is deposited in PubMed Central: PMC13095468, the notice for the Benoist paper, which serves its complete text. It records that following an investigation by Washington State University, Figures 1B and 2A/C and data in the subsequent erratum submission were found to contain falsified or fabricated data, and it names two of the authors as solely responsible. The notices for Kawas 2011 and Kawas 2012, PMIDs 40312094 and 40312092, have no PMC deposit and their wording remains unread. The fact and date of all three retractions are verified from PubMed publication types and Crossref metadata; the stated cause is recorded here for one notice and untraced for the other two. The reason field carried in the Retraction Watch rows is a database classification attached to the 2021 concerns, not the wording of any notice.

The paper still under an expression of concern remains readable in full. Its PubMed Central deposit, PMC3533412, serves the complete text and a PDF beneath a banner stating that an expression of concern has been published and pointing at 378(3):313. That deposit therefore carried its status and its full text together, which the three retracted records did not. Withdrawal from the citable record is not deletion, and a flagged paper that stays legible while carrying its own status is more useful than one that disappears. None of the four papers was removed from the record; what the checks above show is that saved copies and some databases continue to carry them without the flag attached.