The key change: this regimen was already approved for patients who can’t tolerate cisplatin. It’s now approved for all MIBC patients who are candidates for cystectomy — including the cisplatin-eligible population that has long defaulted to gemcitabine + cisplatin chemo.

The expansion rests on KEYNOTE-B15/EV-304, which randomized 808 cisplatin-eligible patients head-to-head against neoadjuvant gem/cis. Perioperative pembro + EV significantly improved both event-free survival (HR 0.53) and overall survival (HR 0.65).

Bottom line: for the first time, an immunotherapy + ADC combination beats cisplatin chemo in the curative-intent setting — and it’s now on the table for essentially every surgical MIBC patient.

🔗 FDA: https://www.fda.gov/drugs/resources-information-approved-drugs/fda-approves-pembrolizumab-or-pembrolizumab-and-berahyaluronidase-alfa-pmph-each-enfortumab-vedotin

What are paper-mills and what is wrong with them? Well, most of the time, papers coming out of these “labs” and authors just aren’t good quality. And “not good quality” can mean a lot of things: not reproducible (small sample sizes, underpowered studies), asking and answering a question nobody cares about, boring articles that fish for false-positive results at a pathway or a gene family with zero context or insight, sloppy or superficial interpretation of results, recycling the same simplified statistical toolkit across every paper regardless of whether it fits the data, and sometimes outright fabricated data and figures.

The manuscript does not specify what % of paper-mill-like papers are coming from not-confirmed paper-mills. I would regard this as more like a continuous scale. Some labs/authors definitely are or are not mass-producing bad science, and some are somewhere in the middle. But I’m sure that most of the labs and research groups are trying and doing their best to do good science.

I haven’t seen a recent study specifically scoring paper-mill output on reproducibility, but I wouldn’t be surprised if it came in low. And let’s not forget that the lack or limited reproducibility even in well-intended, non-paper-mill manuscripts is a well-known phenomenon (a large fraction of medical and bio research doesn’t reproduce).

There are scientists in the world who publish more than 60 papers a year. (I’m fairly confident most scientists don’t even read 60 papers thoroughly in a year.) Some of these people are genuine super-collaborators, embedded in working groups, consortia, multi-site clinical trials. Others show patterns that look a lot more like a paper mill than a scientifically productive career. My selfish definition of a scientifically productive career is regularly publishing (you can argue that Watson/Crick only published one paper throughout their careers, but for better or worse those were different times).

Paper mills are a business. They exist because there’s money in it, and because institutions create the demand. And in my mind, universities share the blame here: chasing rankings, raising publication requirements, and expecting more output without giving people more research time. You can’t demand more papers and more rigor at the same time without paying for either.

The preprint itself

The authors analyzed 33,159 papers published between 2012 and 2023 across 20 high-impact molecular-oncology journals and found that papers suspected of paper-mill origin (according to their previously published method) picked up 50–100% more citations than papers not classified as paper-mill-like did in the one-to-three years after publication, while simultaneously getting fewer actual reads and online accesses. They received fewer recorded online accesses, yet they were cited more than other papers.

The Matthew effect strikes again, you might think — the rich get richer, well-cited papers get cited more simply because they’re already visible. (Worth keeping in mind, too: there’s research suggesting LLM-assisted writing and search may end up amplifying the Matthew effect further, making the already-visible even more visible.) But according to the preprint’s authors, that’s not quite what’s happening here. What they’re describing isn’t an emergent snowball effect — it’s coordinated citation manipulation inflating the impact metrics of journals in molecular oncology, and of authors (the number of highly cited authors is also part of some university rankings).

It used to be relatively easy to spot the weaker end of the publications just by looking at which journal a paper was accepted in. According to this preprint, that heuristic no longer holds. The findings raise the possibility that coordinated citation practices are inflating journal-level metrics. The impact factor, in its simplest form, is the average number of citations papers in a journal received in a year, divided by how many papers that journal published in the two preceding years. It’s a number that used to correlate reasonably well with a journal’s actual standing. Increasingly, it doesn’t.

Figure 2: demonstrating the yearly percentages of flagged papers across leading molecular cancer research journals

And it’s only going to get worse. The scientific publishing landscape has genuinely exploded in the last decade — new journals, new “special issues,” new venues appearing faster than anyone can vet them. To put rough numbers on it: something like 10–12% of the journals active today didn’t exist in 2020, 20–25% didn’t exist in 2015, and 35–40% didn’t exist in 2010.

There are many types of bad science — falsified data, results quietly left out because they didn’t fit the story, and plenty more that has nothing to do with paper mills at all. That’s the part worth remembering: not every bad paper is a paper-mill product. Most of it is just bad science, produced by real labs, under real pressure. Which is exactly why reading everything with a critical eye matters — ask what question is actually being answered, ask whether the method fits the claim. Only a tiny fraction of bad papers ever get corrected or retracted, so the burden mostly falls on the reader.

If you found this interesting PLEASE share it!

Share

Technical note

The Nature News piece covering this preprint has a small error worth flagging: it states that the team “used an artificial-intelligence tool called BERT that they had developed previously.” They didn’t develop BERT. BERT (Bidirectional Encoder Representations from Transformers) is a foundational language model released by Google in 2018, and it underpins a huge chunk of modern NLP (natural language processing), the broader field concerned with extracting reliable information from text and classifying it. What Scancar, Byrne, Causeur, and Barnett actually built is a BERT-based classifier, fine-tuned to flag paper-mill characteristics in cancer-research titles and abstracts — described in their earlier methodological paper in the BMJ (see the link below).

Leave a comment

Source: Nature News, 2026-07-01 — Paper mill cancer studies get double the number of citations as genuine papers (reporting on an underlying bioRxiv preprint, not yet peer-reviewed)

The underlying classifier is described in: Scancar B, Byrne JA, Causeur D, Barnett AG. Machine learning based screening of potential paper mill publications in cancer research: methodological and cross-sectional study. BMJ 2026;392:e087581.

Additional Resources
For Better Science
Retraction Watch
Science Integrity Digest
PubPeer
Csaba Szabo, Unreliable: Bias, Fraud, and the Reproducibility Crisis in Biomedical Research

Subscribe now