Report 050 · Lab Science
A testing lab published what came through its door in 2024 and 2025: of the human cell lines sent in for identity checks, a few percent were not the cell line on the vial. That is the number the coverage will quote. The sentence the authors wrote about their own data is the one worth reading, and it says the true rate cannot be this low.
Every measurement rests on a chain of things you did not check personally. In my own work, spectroscopy, that chain is short and visible: the instrument gets calibrated against a known standard, and if the calibration drifts you find out, because the standard stops reading like the standard. I have written before about what happens in biology when a core reagent never gets that treatment, when more than half of a set of commercial antibodies failed a proper control and the papers had already been written.
This is the same category of problem, one link further up the chain, and it is worse. Not the instrument, not the reagent. The sample. The cells themselves.
On July 9, 2026, Eleanor Ralston and Charlotte Haskayne published something unusually useful in Frontiers in Cell and Developmental Biology: not a study of the problem, but a service provider's account of its own caseload. They work with Biofortuna and NorthGene in the UK, a lab that does cell line authentication for a living, and they reported what actually arrived. In their words, "One thousand, eight hundred and ninety-three samples were sent to NorthGene™ for cell line STR profiling. One thousand, three hundred and twenty-eight of these samples were immortalised human cell lines sent for authentication."
The method is short tandem repeat profiling, the same class of technique used in forensic identification. Every human cell line has a genetic fingerprint at a standard panel of sites. You read the fingerprint and compare it to the reference profile for the line it claims to be. It is cheap, it is fast, and journals, funders and regulators increasingly ask for it.
The results, from the abstract: "4.7% of lines misidentified and 1.8% contaminated in 2024, and 2.4% misidentified with 1.6% contaminated in 2025." There is a third bucket that matters as much. A further 3.6% in 2024 and 2.4% in 2025 came back inconclusive, meaning the match to the expected profile fell "between 59% and 80%" on the standard scoring algorithm, neither a clean match nor a clean miss. When they took those inconclusive results apart with mixture analysis, 50% of the 2024 cases and 66.7% of the 2025 cases turned out to contain contamination. So the ambiguous pile was not noise. It was mostly the same problem, partially hidden.
Read literally, that is a few percent of cell lines being something other than what the researcher believes. In a stock of cells that will be used to generate every figure in a paper, a few percent is not a rounding error. But the literal reading is not the right one, and the authors say so themselves.
Here is the sentence, from the paper, that should govern how anyone cites this work:
"The numbers above contain an unknown but inherent bias as they come from organisations who routinely send the cells for STR-based authentication. A significant number of laboratories still do not undertake this practice and are more likely to only check for microbial contamination."
I want to be clear about how much respect that deserves. A commercial lab had every incentive to present its numbers as a clean prevalence estimate and instead told you exactly why they are not one. But sit with what it means. The population being measured is labs that voluntarily pay to check. That is a self-selected group of the most careful people in the field, the ones with a standard operating procedure and a budget line for verification. And within that group, several percent of samples still came back wrong.
This is a sampling problem, and it runs in only one direction. You cannot measure the error rate of the labs that never test, because not testing is precisely what makes them invisible to the measurement. So the honest way to state the finding is not "about 3% of cell lines are misidentified." It is: among the labs that check, a few percent are wrong, and that is a floor. The unchecked population is, by construction, less careful, and its rate cannot reasonably be lower.
How much higher is a genuinely open question, and I will not manufacture a number. What I can tell you is that estimates from the wider literature run above the tested rate. In a July 2025 letter in Research Integrity and Peer Review, Ralf Weiskirchen of RWTH Aachen University Hospital notes that "another, more recent estimate suggests that 8.6% of all cell lines in use are misidentified." That is an estimate he cites, not a measurement he made, and it should be held as such. It is in the direction the selection bias predicts.
A bad instrument reading is recoverable. You find the calibration error, you apply the correction, you re-analyze. A misidentified cell line is a different kind of damage, because the experiment was performed correctly on the wrong subject. The numbers are real. The statistics are valid. The conclusion is about a different cell type than the one in the title, and no amount of reanalysis rescues it.
Then it propagates. The International Cell Line Authentication Committee keeps a register of lines known to be misidentified, and as of version 14, released February 15, 2026, it lists 608 of them. The detail in that register I find hardest to shake: 560 of those lines have no known authentic stock anywhere. There is no correct version to go back to. The original was lost, and what circulates under the name is something else, usually something more vigorous that outgrew it in a shared incubator. HeLa alone accounts for 145 entries.
The published record has already absorbed this. Weiskirchen reports that "a comprehensive search of the PubMed database identified almost 6,000 publications using the misidentified cell lines QGY-7703, BGC-823, BEL-7402, L-02, and WRL-68," five lines out of six hundred. He cites a broader estimate that "32,755 studies have used misidentified cells, which in turn have been cited in approximately half a million subsequent publications." Again: an estimate, cited, not a headcount. Treat the order of magnitude, not the digits.
And the checkpoint that was supposed to catch it did not. Reviewing four cases where misidentification was raised with journals after publication, Weiskirchen describes outcomes ranging from a promptly published commentary to a journal that rejected the letter on scope grounds and stopped responding. On the review stage itself, his observation is that reviewers "did not request or investigate information about the cell line used, which would have been essential to prevent the publication of inaccurate research data." Nobody asked. That is the whole failure, and it is the same failure as with the antibodies: a shared assumption that provenance is somebody else's job.
One more finding from the 2026 paper deserves its own line, because it is about the cells people trust most. Of the primary cell lines submitted in that period, none met the ASN 0002 standard for full traceability. Not a low percentage. None.
The fix here is not exotic, which is what makes the situation frustrating rather than tragic. STR profiling exists, it is inexpensive relative to any experiment worth publishing, and reference profiles are public for most established lines. The paper's own recommendation is institutional: authentication should be the institution's responsibility, done before culturing rather than after a result looks strange.
If you read papers rather than write them, this gives you one specific question that is easy to check and rarely answered: does the methods section say the line was authenticated, when, and by what method? A paper that names its cell line and says nothing about verifying it is not necessarily wrong. It is unverified, which is a different claim from correct, and you are entitled to know which one you are reading. That is the same move I have argued for with a beautiful micrograph and with a dramatic behavioral result: separate what was measured from what was assumed.
Two honest limits on everything above. The Frontiers figures are one commercial provider's intake over two years, not a random sample of the world's labs, and the drop from 4.7% to 2.4% between 2024 and 2025 is two data points from one lab's caseload, not a trend anyone should announce. The 8.6% and the 32,755 are estimates cited in a letter, which is a weaker form of evidence than a direct count. None of that changes the structural conclusion, because the structural conclusion does not depend on the exact figure.
Identity is a measurement. It is just the one nobody thinks of as a measurement, because it arrives written on the side of a vial in someone else's handwriting, and a label is the most persuasive uncontrolled variable in science. Everything downstream, every instrument you did calibrate, every control you did run, is conditional on it being right.
The most valuable thing in this year's data is not the percentage. It is that the people who produced it told you their sample was self-selected. That single sentence turns a prevalence claim into a floor, and a floor is a far more useful thing to know. When a number comes with an honest account of who was and was not counted, you can reason with it. When it does not, you are reading a measurement with an unstated boundary, and I have written about where that leads in a field with a great deal more money in it.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.