Report 038 · Lab Science
A team ran the experiment almost nobody runs: buy the commercial antibodies for a protein, then test them against cells engineered to have no such protein at all. If the antibody still lights up, it was never detecting what the label said. More than half of 614 antibodies failed. A measurement scientist on the most-used uncalibrated instrument in biology.
I do measurement science. In spectroscopy, nobody would let you publish a peak assignment without telling them how the instrument was calibrated, what the blank looked like, and what a known standard did on the same day. That is not bureaucracy. A measurement is a claim about reality made through a device, and the claim is only as good as your evidence that the device does what you think.
Cell and molecular biology runs on an instrument that mostly skips this. It is called an antibody, and it is the single most common way biologists ask "is this protein here, and how much of it?" It is used to stain tissue, to pull a protein out of a mixture, to produce the bands on a Western blot that anchor a huge fraction of the biomedical literature. And for most of its history, the evidence that a given antibody actually binds the protein on its label has been, essentially, the vendor's word.
There is a clean way to test this, and it has been available for years. Take a cell line. Use gene editing to knock out the gene for your target protein, so the protein simply is not there. Now run your antibody on both the normal cells and the knockout cells, side by side. The normal cells should give a signal. The knockout cells should give nothing. That is it. That is the control.
If the antibody produces a band or a stain in cells that cannot make the protein, then whatever it is detecting, it is not the thing you are naming in your paper. This is the biological equivalent of running your blank, and it is the closest thing the field has to a gold standard.
A consortium called YCharOS, working with antibody manufacturers and doing all of it in the open, has been running this test at scale. Their results, published in eLife in 2024, are worth reading slowly. They "scaled a standardized characterization approach using parental and knockout cell lines to assess the performance of 614 commercial antibodies for 65 neuroscience-related proteins." The headline finding: "more than 50% of all antibodies failed in one or more applications."
More than 50% of all antibodies failed in one or more applications.
Two things deserve emphasis, in both directions. First, this is not "half of all antibodies are garbage in every use." The failure is application-specific. An antibody can be perfectly good on a Western blot and useless for immunofluorescence, because those are physically different measurements: one detects a denatured, size-separated protein on a membrane, the other detects a folded protein sitting in a fixed cell among everything else. Passing one says little about the other, which is precisely why "validated" with no application named is not information.
Second, the news is not uniformly bad. The same study found that "~50–75% of the protein set was covered by at least one high-performing antibody, depending on application." For most targets a good reagent exists. The problem is that it is sitting on a catalog page next to several bad ones, at a similar price, with similar marketing.
Failed reagents would be a manageable annoyance if they were caught before publication. The companion eLife Science Forum piece, written by a large group including researchers, funders, and journal representatives, reports what actually happens. Across the targets they examined, "an average of ~12 publications per protein target included data from an antibody that failed to recognize the relevant target protein."
Twelve papers per protein, on average, built on a reagent that does not detect the protein in question. Those papers are in the literature now. They are cited. Some of them have shaped which hypotheses got funded and which drug targets looked promising. Nobody committed fraud. Everyone ran the assay the way they were taught, using a product sold as specific for the target.
The same piece puts the scale plainly: "~50% of commercial antibodies fail to meet even basic standards for characterization, and this problem is thought to result in financial losses of $0.4–1.8 billion per year in the United States alone."
That figure sits inside a larger, older one. The 2015 PLOS Biology analysis by Freedman, Cockburn, and Simcoe estimated roughly $28 billion per year spent in the United States on preclinical research that is not reproducible, and attributed about 26% of that to "biological reagents and reference materials," one of the two largest categories alongside study design. The antibody problem is not a curiosity at the edge of the reproducibility crisis. It is a substantial fraction of its mass.
It helps to understand what an antibody actually is, because the failure mode follows from the biology.
A polyclonal antibody is not one molecule. It is the serum response of an animal immunized with your target, meaning a mixed population of antibodies recognizing many different features, some of which are on your protein and some of which are on whatever else resembles it. That mixture is finite and unrepeatable: when the animal is gone, the reagent is gone, and the replacement lot is a different mixture with different behavior. A monoclonal is a single clone, so it is consistent, but consistency is not specificity. A monoclonal can be beautifully reproducible at binding the wrong thing.
Recombinant antibodies, where the sequence is known and the reagent is expressed from that sequence, fix the reproducibility half of the problem, and the YCharOS data suggest they also do better on the specificity half: "recombinant antibodies performed better than monoclonal or polyclonal antibodies." In their Western blot results, 67% of recombinant antibodies immunodetected their target, against 41% of monoclonals and 27% of polyclonals.
Layer onto that an economics problem. Rigorous knockout-based characterization costs real money and can only lose the vendor sales, since a failed antibody is one you cannot sell. Meanwhile the buyer, a graduate student with a deadline, has no practical way to distinguish a well-characterized reagent from a poorly characterized one at the moment of purchase. That is a market where the diligence does not get rewarded unless someone external does it and publishes the answer, which is exactly the gap YCharOS was built to fill.
If you use antibodies, four habits carry most of the value:
Check whether someone already tested it. YCharOS results are published openly, and antibody-registry identifiers (RRIDs) let you find the exact reagent rather than a vague description. Look before you buy, not after the figure fails to replicate.
Demand the control that matters. A knockout or knockdown line is the strong one. Where that is impossible, orthogonal confirmation, measuring the same protein by an independent method such as mass spectrometry, is the next best thing. As the Science Forum authors put it, "the characterization of antibodies is always further improved when combined with other approaches."
Treat "validated" as an unfinished sentence. Validated for which application, in which cell type, at what dilution, against which control? A vendor image of a single clean band is a photograph, not a characterization.
Prefer recombinant where one exists. You get a defined sequence, a reagent that still exists in five years, and on the current evidence, better odds.
And if you are a reader rather than a bench scientist, there is one thing worth carrying away. When you see a paper reporting that a protein is elevated in a disease, or localizes to some compartment, ask quietly what that measurement was made with. Frequently the whole edifice rests on an antibody, and frequently nobody checked the antibody.
Every field has a favorite instrument it forgets is an instrument. In spectroscopy we have Raman, which is used as reflexive proof in materials papers where it proves very little. In microscopy it is super-resolution reconstruction, which can manufacture detail that was never there. In cell biology it is the antibody. The pattern is identical each time: a technique becomes routine, routine becomes invisible, and invisible means uncalibrated.
The encouraging part is that this one has a fix, it is cheap relative to the waste, and it is already running in the open. The uncomfortable part is the twelve papers per protein that are already published. Nothing recalls those. The only available move is to make the next twelve better, which starts with treating a bottle of antibody the way you would treat any other measuring device: as a thing that must earn your trust with a control, every time, for the specific measurement you are about to make.
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.