Science is publishing papers on proteins that do not exist. That is why it is no longer enough to trust studies

Scientific publishing has become a system in which it is possible to publish a biologically impossible experiment, have it pass peer review, and then find it, years later, cited as evidence of the existence of a phenomenon. By the time someone discovers the error, the result has already had time to contribute to the development of new hypotheses, guide further experiments and perhaps even justify funding. The case, made public on 21 October by independent researcher Sholto David, shows just how far this process has gone, as it concerns experiments that should have been recognised as flawed on the basis of biological knowledge that has been available for decades.

David has compiled 230 reports of protein expression that is incompatible with the known genetic characteristics of the cells studied, spread across 221 publications. The principle behind his investigation is very simple. When a gene has been completely removed from a cell’s genome, that cell cannot produce the corresponding protein. If a scientific article features a photograph in which the protein appears nonetheless, something in the experiment must necessarily be wrong. Yet those photographs were published, often accompanied by elaborate interpretations of the biological mechanisms they were supposed to demonstrate.

Let us consider a case dating back to 2008. MCF-7 cells, derived from a breast tumour and widely used in cancer research, have a 47-base deletion in the caspase-3 gene, a protein involved in programmed cell death. This alteration prevents the production of the normal protein, a fact documented as far back as 1998 in a paper published in the Journal of Biological Chemistry. However, an article published in the International Journal of Oncology in 2008 demonstrated, using a technique called Western blot, the presence of caspase-3 in those cells. Western blotting enables the detection of a protein using an antibody and produces a dark band at its location. In the case of MCF-7 cells, that band should not have been present.

One might think this was an isolated error. In 2015, a paper published in Anti-Cancer Drugs, focusing on a natural substance called zerumbone, described the effects of the treatment on the STAT3 protein in PC-3 prostate cancer cells. The problem is that these cells have a deletion encompassing the entire STAT3 gene, a fact already documented in 2003 in Oncogene. Nevertheless, the authors presented a Western blot indicating its presence, basing part of their interpretation of the observed pharmacological effect on that signal.

Clearly, we do not need generative artificial intelligence to produce such results. The examples from 2008 and 2015 predate by a long way the widespread use of current systems capable of generating plausible experimental images. The phenomenon has deeper roots and includes possible exchanges of cell lines, antibodies that recognise proteins other than those stated, experimental errors and the manipulation of results. David’s compilation identifies contradictions that require explanation; it would be incorrect to automatically attribute each one to deliberate fraud.

The result remains devastating. Peer review failed to identify contradictions that could have been verified by consulting already available genetic information. In another example discussed by David, an antibody used to recognise the BAX protein – cited in over 1,400 publications – was only identified as problematic in 2024, through a study published in Cell Death & Disease. This means that for years, the same reagent may have underpinned a vast number of scientific findings whose interpretation is now in doubt.

The problem becomes even more serious when such findings find their way into literature reviews. David demonstrates how studies on STAT3 signalling in PC-3 cells – which are incompatible with their genetic makeup – have been cited in review articles. At that point, the erroneous information begins to appear as though it has been confirmed by the convergence of various publications. Artificial intelligence systems that analyse the literature can reproduce and further amplify it. The proliferation of citations thus lends an appearance of validity to what was incorrect from the outset.

And yet David has identified only those errors that can be recognised thanks to a previously known biological incompatibility. How many false results remain undetected because they describe events that are biologically possible but have never actually occurred? How many experiments that were never carried out are indistinguishable from genuine experiments once they have been transformed into figures in an article? David’s sample, compiled through targeted research, does not allow us to estimate the percentage of erroneous publications. It does, however, demonstrate that even elementary contradictions can survive the scientific scrutiny system for years.

The success of fraud also depends on an asymmetry that will continue to worsen. To produce a false result, one need only alter a photograph, invent a set of numbers or present a non-existent experiment as one’s own. Recognising this requires specific expertise, access to the original data and, often, an investigation that takes weeks of work. When a technique capable of detecting manipulation is developed, those seeking to deceive can learn to circumvent it. It is an evolutionary race in which methods of falsification are selected precisely because of their ability to bypass the available checks. Artificial intelligence accelerates this process, reducing the cost of producing seemingly credible results.

Continuing to invest primarily in detecting fraud after publication means accepting that we are chasing an adversary who can constantly change their strategy. We need to intervene at an earlier stage, when the scientific data is being produced.

The necessary technology exists. A modern laboratory instrument generates digital information that can be automatically recorded, linked to the identity of the instrument and the experiment, time-stamped and protected using cryptographic procedures. A digital signature can certify the document’s origin, whilst a specific mathematical fingerprint, known as a hash, makes it possible to verify whether even a single element of the file has been subsequently altered. By recording these fingerprints in a distributed ledger – a blockchain – it becomes possible to construct a verifiable history of the data’s creation and any subsequent changes.

Technologies of this kind have already been studied and tested in clinical research. In 2022, Scrybe was unveiled – a system that uses a permissioned blockchain to record the provenance of clinical data, with a prototype integrated into the REDCap platform, which is widely used to collect trial data. The system is described in an article published in Digital Threats: Research and Practice. A paper published in December 2025 in Frontiers in Medicine described the implementation of an architecture based on Hyperledger Fabric, verified using data from a pilot clinical trial, to ensure the traceability of operations and information. These applications are still limited, but they demonstrate that the key technological tools are available.

Of course, certifying the origin of a piece of data does not automatically prove that the experiment was conducted correctly. Even a genuine instrument can measure the wrong sample, and a defective antibody can generate a perfectly genuine signal. The cryptographic guarantee concerns the provenance and integrity of the record, whilst the biological quality of the experiment requires appropriate controls. In the case of the Western blots examined by David, the genetic identification of the cell lines and the validation of the antibodies therefore remain essential. Data certification would become truly effective by linking this information to the instrumental results from the moment they are acquired.

We can therefore envisage a laboratory in which the digital original of every experiment is automatically recorded in an independent archive, together with the information needed to interpret it. A researcher could freely process it, improve the visualisation of an image, carry out statistical analyses, correct errors and repeat calculations. Every step would remain verifiable, as the initial data would be preserved and the transformations documented. The greatest advantage would be that it would make it much more difficult to retroactively replace unwelcome results with more convenient ones, without leaving a trace.

To implement a similar system, tools are required that can stamp data directly upon acquisition, reliably identify operators and samples, and provide independent archives and verification procedures. Blockchain would be one of the technologies that could be used to guarantee the immutability of records, without having to store images or entire datasets on it. Cryptography already makes it possible to separate data storage from proof of its integrity. Regulatory authorities, moreover, have long required original records and systems for tracking changes – so-called audit trails – in numerous pharmaceutical and clinical contexts. Extending these principles to routine scientific research would be a challenging change, but one that is perfectly compatible with current technological capabilities.

Such a transformation, however, requires us to address an even more fundamental question: what role should scientific publication play?

Today, an article published in a journal continues to represent, in practice, the primary unit of certified knowledge. It is the item that is cited, used in academic assessments, included in systematic reviews and analysed by artificial intelligence. Yet the article is essentially the authors’ account of their own experiments. The data presented are those they have selected and organised to support their conclusions. The publishing system verifies this account primarily through the judgement of other scientists, who are rarely provided with the tools necessary to reconstruct the entire experimental process.

We should begin to regard publication as the moment when interpretations of research are discussed, thereby granting experimental data an autonomous and verifiable existence. The fundamental heritage of science would thus consist of large, interoperable archives of original data, accompanied by information on their production, quality controls and available certifications. Verification of provenance and that of scientific validity would be explicitly distinguished, and a result could be re-evaluated without relying exclusively on the article in which it was presented.

A significant proportion of the results could also be subjected to automated checks even before they enter the scientific literature. It would be possible to compare the declared genetic identity of a cell line with the protein that an experiment claims to measure, to check the reagents used, and to verify the consistency between original data and published analyses. David’s work demonstrates precisely that information already available in various archives can be combined to identify contradictions that have escaped human detection.

The benefit of having certified scientific databases would become even greater in the age of artificial intelligence. These systems acquire knowledge from a body of literature that contains a mix of reliable findings, errors and fabrications. They are capable of rapidly processing volumes of literature that no scientist could possibly read, but the quality of their responses also depends on the quality of the information used. The case of the non-existent proteins demonstrates just how easily an error can be reproduced from one publication to another, to the point where it is assimilated by the automated systems we use to navigate our research.

The prospect of further growth in scientific output, fuelled by the automated generation of texts and images, makes it absolutely essential to distinguish between what has actually been measured and what someone has merely claimed to have measured.

For this reason, universities, research organisations, funding bodies and publishers should agree on common standards for certifying data at source and support public infrastructures capable of preserving it in the long term. Research funding should include the production of verifiable records, their preservation and the availability of the metadata essential for verifying them, with the necessary safeguards for confidential data. Journals could require a documented provenance of experimental evidence, clearly indicating which controls were carried out.

Such a transformation would require investment and a fundamental shift in academic practices. It would also make it possible to allocate an increasing proportion of scientific resources to the production of reliable knowledge, rather than expending them in an attempt to identify false information retrospectively.

Sholto David’s work should raise concerns that go far beyond the 221 articles identified. For at least twenty years, findings incompatible with basic biological knowledge have been published, whilst we continue to regard publication as the primary guarantee of a discovery’s reliability. Meanwhile, we have developed technologies capable of verifiably recording the origin of information, and today we have IT systems capable of automatically comparing vast amounts of data.

Continuing to build scientific knowledge without demanding guarantees regarding the provenance of the evidence underpinning it is an increasingly untenable choice. If we allow the production of fakes to grow faster than our ability to recognise them, we risk gradually losing the very ability to distinguish accumulated knowledge from its imitations. The time has come to change the rules, starting with what a tool has actually recorded, before anyone turns it into a publication.