My own take from the Abcam's catalogue.

Some investigations are lonely. This one turned out to be the opposite, and that is the most encouraging thing about it.

Let me start with the timeline, because it matters. On 17 May 2026, Sholto David spotted a single suspicious image in Thermo Fisher's antibody catalogue: a western blot with bands whose shape recurred beyond reasonable chance. He thought it was probably an isolated case, but within a handful of days, he and Reese Richardson had documented more than a hundred problematic images in this same Thermo Fisher's catalogue. By early June the count reached over 450, and another isolated example had surfaced in the catalogue of another antibody vendor: Abcam. The story received a significant press coverage, for example by Nature, and it landed hard, because catalogue images are a marketing evidence a company offers to convince researchers its product works.

I read that early reporting and had a slightly uncomfortable thought. If one major vendor's catalogue contained that many questionable images, what was the prior probability that it was the only one? Vendors are competitors in the same market, they have the same incentives, use the same off-the-shelf software to process images. Why would Thermo Fisher be unique?

So, during that relatively quiet interval in early summer, after the first Thermo Fisher revelations but before anyone knew the true scale, I started doing a slow, manual verification of a different catalogue: Abcam's. Below is what I found, how I arrived at it, what my method can and cannot establish, and how my (admitted) small effort turned out to converge with a much larger one.

I want to be clear from the start: nothing here is proof of misconduct by anyone, and my purpose is not to pillory a company. Instead, I wish to describe a pattern, ask for clarification, publish the response provided by the vendor, and reflect on what this story means for how researchers buy and trust reagents.

What I did, and why my method is a bit different

Abcam's catalogue is gigantic, and manual forensic inspection does not scale to hundreds of thousands of antibody references. I am one person alone without an automated pipeline, so I sampled. A good old fashioned manual sampling. I selected catalogue pages at random and worked through them by hand, examining the validation images, mostly western blots, linked to each antibody. In total I inspected roughly 400 antibodies. Where an image was too small, too compressed, or too low in contrast for meaningful analysis, I excluded it. The rest I ran through Forensically, a free, online photo-forensics toolkit, using four of its modules: Error Level Analysis (ELA), which highlights regions whose compression history differs from surroundings; noise analysis, which isolates the high-frequency noise residual; luminance gradient, which shows edges of brightness changes invisible to the naked eye; and Principal Component Analysis (PCA) that re-projects the image onto its principal components to surface subtle local inconsistencies.

This is worth dwelling on, because it is where my approach differs from the one the larger investigation used. Richardson and colleagues built their case largely on duplication and pattern-matching (the same well-identifiable band appearing twice, dozens or thousands of separate blots sharing one visually identical background, cloned patches of noise), sometimes revealed by adjusting black and white levels. Similarities are powerful evidence because duplication is hardly defendable (close to unarguable) since the same random patterns do not recur by chance. My approach was different because I was not primarily looking for repeated images, I was looking at the internal forensic signatures of single images, asking whether the compression and noise behaviour within one blot was consistent with a single, photograph acquired honestly. Two different methods, aimed at the same question from different angles.

Across my sample I flagged about 15 antibodies whose images showed patterns I considered suspicious enough to motivate a second inspection. For each, I saved the full set of outputs (original alongside ELA, noise, gradient and PCA).

Two recurring patterns

The images I flagged mostly fell into two families.

The first class consists in localised squares or rectangles around one or more bands that stands out from the rest of the ELA or noise image. In a blot acquired and saved as a single coherent image, the error level and noise should be roughly homogeneous across the global image. A rectangular region of different error level following no biological boundary is the kind of footprint left when image content is pasted in from another image and the composite is saved (once again, it is consistent with a copy-paste scenario, but is not, by itself, proof of one).

The second category is "striping": vertical lanes with systematically different error, noise or PCA signatures from their neighbours, with discontinuities falling clearly along lane boundaries. A non-naive reading is that the lanes never shared a gel, that proteins were run separately, imaged separately, and assembled side by side into a single presentable figure. Splicing is not automatically fraudulent, and may be legitimate when the figure is transparently described as composite. The problem arises when a spliced image is presented as one contiguous blot, because the reader then infers things about molecular weight, and relative intensity that the actual experiment never supports.

At higher magnification:

The honest limits, which turned out to be everyone's limits

ELA, noise and PCA are compression- and processing-forensics, and as such they are notoriously confounded. On the Forensically website the tool's own documentation says that error level analysis is sensitive to resizing, recompression, adjustment of contrast, and the simple fact that a sharp band on a clean background stands out from a flat region. Catalogue images are routinely downscaled, compressed more than once, and cleaned for web display, and every one of those legitimate steps can sometimes produce a rectangular ELA halo or a lane-to-lane difference that looks like manipulation but is not.

In other words, my fifteen flags are hypotheses, smoke alarms worth investigating but possibly set off by toast. I have an unknown false-positive rate, no access to the original uncompressed images, and a sample that is not cleanly random (for example: excluding low-quality images is itself a selection effect).

Richardson's investigation last August had the same caveats. That convergence is what good-faith sleuthing looks like when several people do it independently.

A bigger picture than my 400 antibodies

In late August, Richardson published the result of the collective effort, and it dwarfed my little survey. The shared repository now documents more than 18,000 problematic images across some 17,000 antibody products sold by 15 vendors (Thermo Fisher, Abcam, Novus Biologicals, ProteoGenix, OriGene, MilliporeSigma, LSBio, Bioss, Boster Bio, G-Biosciences, GeneTex, HUABIO, Antibodies.com, Abnova and Santa Cruz Biotechnology). The findings were reported across Nature, Science, The Scientist, C&EN and GenomeWeb. It has become a story about an entire industry.

Two findings from that work reframe my own results, and honesty requires me to say so. First, many vendors turn out to use the same images as one another, strong evidence of private labelling, where one manufacturer produces an antibody and its validation data, then sells it to multiple companies who market it under their own brands. Second, several vendors, Abcam among them, have publicly noted that their portfolios include externally-sourced products. Both facts bear directly on my "striping" pattern: some of what I flagged may not have originated at Abcam at all, but arrived as third-party validation data that Abcam republished. That does not make the images fine, it relocates the question. But the question remains.

For the record, the count attributed to Abcam in the collective repository is 185 images across 115 products, arrived at by a completely different method than mine. I find that convergence quite persuasive.

I should also add, on a human note, that when I began noticing suspicious patterns, I have contacted Reese Richardson and exchanged generous and collegial messages with him. That matters to me. This kind of work is often unpaid, professionally unrewarded, and occasionally risky, and the thing that sustains it is worldwide community of people who check each other's work, share methods, and care about the reliability of the scientific record. My contribution here is a tiny tile in a large mosaic, and I am glad the mosaic exists.

Abcam's response

I strongly believe in right of reply. So I contacted Abcam with my observations and offered to share my documented images before writing this blog post. They recently sent me a statement. In fairness and transparency, here it is, signed by Abcam's spokesperson:

In substance, Abcam says every validation image, whether generated in-house or supplied by a partner, is reviewed by Abcam scientists before publication for scientific soundness; that their review methods are evolving to include image-integrity tools that were not previously available; that they review and remove older images that no longer meet current standards; that they participate, since 2020, in the independent YCharOS antibody-characterisation programme and openly publish data on products that do not perform in given applications; and that when they identify a genuine problem (they cite a CD9 antibody withdrawn after a cross-reactivity issue) they do not stay inactive, for example by contacting affected authors.

Some of this is to their credit, true. I will not pretend otherwise. Participating in YCharOS is a real commitment, and publishing negative validation results is more transparency than many vendors offer. But I have to be equally honest about what the response did not do. It did not engage with the specific patterns I described. It did not take up my offer to review the actual images I flagged. It did not commit to providing the original, unprocessed acquisitions that would, in most cases, resolve the ambiguity outright  (which is the one tangible step that would move this from suspicion to certainty). 

What this means for the rest of us

Stepping back from any single vendor, the structural problem is harsh. The antibody market was valued at over US$200 billion last year, with research (non-medical) antibody representing US$4B alone, and a 2023 eLife survey found that more than half of 614 commercial antibodies did not perform as specified. Validation images are the front line of how a buyer decides whether to trust a reagent. If those images are "improved" cosmetically, spliced, or fabricated from a shared template before publication, there is no ground for trust anymore.

Richardson draws a hard economic conclusion from this: proper knockout-based validation costs far more than most antibody products ever earn, so the incentives favour selling a large catalogue of untested reagents over a small catalogue of properly-validated ones. He goes as far as "I think that the biomedical research community would be better served by the creation of publicly-funded antibody vendors whose interest is in producing antibodies that work, not in producing antibodies that sell". I think the diagnosis is sound and worth taking seriously, the boldest version of the prescription is a bigger debate than I can settle here, and reasonable people will disagree about it. But before endorsing the restructuration of the entire industry, I would start with an unavoidable conclusion.

That conclusion is about audits and trust. Other high-stakes industries such as aviation, clinical diagnostics, finance, are subject to routine, independent audit, and no one considers this an insult. The antibody industry, whose products underpin an enormous fraction of biomedical research, I think has to change to adopt similar independent, systematic auditing of catalogue validation data should become a norm. Driven by open-science consortia like YCharOS, by funders, or eventually by regulation, whatever... I don't settle it here. Voluntary self-review is not sufficient on its own.

And until that norm exists, the practical burden falls on the buyers. So my immediate recommendation is not to blindly trust the validation image on the product page, treat it as a marketing claim, not as data. For any antibody your results will genuinely depend on, ask the vendor for the original, unedited validation image and validate the signal yourself in your own system where you can. Use knockout or knockdown controls, and lean on independent resources like YCharOS. 

A closing note

I have written before that visual trust in science is eroding, and that generative AI is dismantling the last fragile safeguards against fabricated images. This episode is the same problem seen from the prism's other end. As we worry about fake images entering the literature, we have paid far less attention to the images on the commercial pages.

What gives me hope is the shape of how this unfolded, with one person noticing one odd image, then a handful of independent sleuths using different methods quietly converging on the same truth. Then, careful journalists, then (slowly) the responses from the vendor companies. My small survey of 400 antibodies is one mini-brick in that wall, there will be many more.

My request to Abcam remains simple and made in good faith: show us the originals. It is the fastest way to turn dozens of question marks into answers and I would be glad to be convinced I was wrong about some of them.

References and sources

Richardson R blog article. 2026, "How much of Thermo Fisher’s antibody data has been manipulated?": https://reeserichardson.blog/2026/05/28/how-much-of-thermo-fishers-antibody-data-has-been-manipulated/

Richardson R 2026 blog article, "At least 15 companies are selling antibodies using faked validation data": https://reeserichardson.blog/2026/08/25/at-least-15-companies-are-selling-antibodies-using-faked-validation-data/

Richardson R, David S et al. 2026, the antibody validation image repository (Zenodo): https://doi.org/10.5281/zenodo.20402475

Garisto D 2026, Nature news coverage of the initial Thermo Fisher findings: https://www.nature.com/articles/d41586-026-01706-2

Forensically, the photo-forensics toolkit used for this analysis: https://29a.ch/photo-forensics/

Ayoubi R et al. 2023, eLife survey of commercial antibody performance: https://doi.org/10.7554/eLife.91645.2

Banner created by Gemini Banana Pro. Text drafted by Claude (a bullet-point plan) and then written by RDG. Forensic analyses performed by RDG using Forensically. Abcam was contacted for comment prior to publication and their response is represented above.