The retraction notice appeared quietly in July 2024: a paper that had linked early-life environment to stress circuit function was being pulled because the authors could no longer determine which mouse substrain they had used. The study, originally published in the Journal of Neuroscience, had reported that a specific stress circuit in the mouse brain—the hypothalamic-pituitary-adrenal (HPA) axis—responded differently to a mild stressor depending on the animal's early-life environment. The finding fit neatly into a growing literature on developmental programming of stress reactivity. But when two independent labs tried to replicate the result using the same protocol, they got opposite answers. One group saw the expected blunted cortisol proxy; the other saw no effect. The discrepancy took nearly two years to trace to its source: a mouse substrain shift that had gone unrecorded. The story of how one unreported substrain change bent a stress hormone circuit paper is a cautionary tale about the hidden variables that live inside the most common research tools.

The $2 Million Mouse That Bent a Stress Circuit

The C57BL/6 mouse is the workhorse of behavioral neuroscience. It accounts for roughly half of all rodent studies in the field, and its genome was the first mouse genome sequenced. But C57BL/6 is not one animal. It is a family of substrains that diverged after the 1950s when breeding colonies were separated. The two most common are C57BL/6J, maintained by the Jackson Laboratory, and C57BL/6N, distributed by Charles River Laboratories and others. For years, researchers treated them as interchangeable. They are not. A 2013 study showed that C57BL/6J and C57BL/6N differ in over 150 single-nucleotide polymorphisms, including a cluster on chromosome 11 that affects the regulation of corticotropin-releasing hormone (CRH). CRH is the master switch of the stress response. In C57BL/6J mice, the Crh promoter is more methylated, leading to lower baseline corticosterone. In C57BL/6N mice, corticosterone runs roughly 0.3–0.8 log fold higher. That difference is large enough to flip the outcome of a stress circuit experiment.

The retracted paper had used mice from a vendor that switched its breeding stock from the J to the N substrain in 2005 without changing the catalog number. The authors, unaware of the switch, reported their animals as C57BL/6. The stress circuit they were studying—the paraventricular nucleus of the hypothalamus projecting to the median eminence—turned out to be exquisitely sensitive to baseline CRH tone. A 15–20 percent shift in firing rate of CRH neurons, as observed between substrains, was enough to change the direction of the experimental effect.

Anna K. Lee, a postdoctoral fellow at the University of California, Berkeley, who led the replication analysis, estimates that the total cost of the original study—including animal breeding, personnel, and imaging time—was roughly $2 million over six years. “That money didn't just vanish,” she says. “It produced a result that looked solid but was actually an artifact of a variable nobody thought to report.”

How a Single Vendor Swap Created Unseen Variance

The vendor swap was not secret. Jackson Laboratory and Charles River Laboratories both publish detailed genetic quality-control reports for their colonies. But those reports are not routinely read by end users, and the substrain designation is not printed on the shipping label. A lab ordering C57BL/6 from a distributor that sources from Charles River will receive C57BL/6N unless it specifically requests the J substrain.

The problem is compounded by the fact that many behavioral neuroscience protocols do not include a step to verify substrain. Genotyping panels for mice often include markers for coat color and common disease genes but not for the polymorphisms that distinguish J from N. A lab could run a full SNP panel and miss the difference because the relevant markers are not on the standard array.

Lee's reanalysis of 12 published HPA axis studies found that substrain explained 40–60 percent of the variance in corticosterone levels—more than any experimental manipulation. “That's a staggering number,” she says. “It means that if you don't control for substrain, you are basically rolling dice on whether your control group is a J or an N, and that alone can determine whether you get a significant result.” The 2021 paper's authors had used a protocol that had worked in C57BL/6J mice but failed in C57BL/6N because the baseline corticosterone was already elevated, masking the effect of the stressor.

Facility managers at several large academic institutions, when contacted by Lee's team, acknowledged that they had refreshed their breeding colonies from Charles River without notifying principal investigators. “It's not malicious,” Lee says. “It's just that the substrain was not considered a relevant variable. Now we know it is.”

The Preprint That Exposed the Hidden Variable

Lee's reanalysis appeared on bioRxiv in early 2023. The preprint, titled “Substrain identity explains a large fraction of variance in murine HPA axis measures,” analyzed both published data and new measurements from a cohort of mice whose substrain was verified by SNP genotyping. The results were stark: studies that reported using C57BL/6 without specifying J or N showed a bimodal distribution of corticosterone values, with peaks corresponding to the two substrains.

The preprint was not immediately accepted for journal publication. Reviewers questioned whether the substrain effect could really be large enough to explain the replication failures. “There was pushback,” Lee recalls. “Some people said, 'We've been using C57BL/6 for 20 years and never had a problem.' But that's exactly the point—they never checked.”

Within months of the preprint, three other groups posted comments identifying additional retractions or corrections that could be traced to substrain issues. One was a 2019 paper on fear conditioning that had used C57BL/6N mice but cited earlier work done with C57BL/6J. Another was a 2020 study of antidepressant efficacy that had reported a null result; reanalysis showed the drug worked in J but not N substrains. The pattern was clear: the hidden variable was everywhere.

Lee's preprint also sparked a broader conversation about metadata reporting in animal studies. “We ask authors to report the strain, but not the substrain,” says Michael Eisen, a biologist at the University of California, Berkeley, who was not involved in the work. “That's like asking for a car's make but not the model year. The difference matters.”

Why Standard Housing Protocols Missed the Drift

The Guide for the Care and Use of Laboratory Animals, the authoritative document that governs animal research in the United States, does not require substrain reporting. It specifies that animals should be identified by strain, sex, and age, but leaves the level of detail to the discretion of the institution. As a result, many animal facilities pool J and N substrains in the same housing rooms, and investigators often assume that all C57BL/6 mice are the same.

Behavioral neuroscientists have long relied on C57BL/6 as a monolithic label. Textbooks describe its behavior as consistent across labs. But the substrain drift has been known to specialists for over a decade. A 2015 study in Genes, Brain and Behavior showed that C57BL/6J and C57BL/6N differ in anxiety-like behavior, locomotion, and startle response. A 2018 paper in eNeuro reported differences in synaptic plasticity in the hippocampus. Yet these findings did not percolate into routine practice.

One reason is that genotyping panels for mice rarely include the markers that distinguish substrains. The most common commercial panel, the Mouse Universal Genotyping Array, covers roughly 10,000 SNPs but omits the chromosome 11 region that differs between J and N. A lab would need to order a custom assay to detect the difference, which costs roughly US$5–8 per sample—a trivial expense compared to the cost of a single experiment, but one that few labs considered necessary.

“It's a blind spot created by the assumption that the vendor's catalog number is a reliable identifier,” says Sarah J. Johnson, a mouse geneticist at the University of North Carolina. “But vendors update their breeding stock, and the catalog number stays the same. The only way to know what you have is to test it yourself.”

The Retraction That Forced a Methods Rethink

The Journal of Neuroscience retracted the 2021 paper in July 2024, after the authors acknowledged that they could not determine which substrain their mice belonged to. Lab notebooks from the period did not record the vendor lot number or the date of colony refresh. The animals had been ordered through a university central facility that pooled orders from multiple vendors. In a public statement, the lead author wrote: “We honestly don't know whether we used J or N. We cannot guarantee the reproducibility of our results.”

The retraction prompted the journal's editorial board to issue new guidelines. As of January 2025, all manuscripts reporting rodent work must include the substrain designation, the vendor, and the lot number of the animals used. The policy also encourages authors to deposit genetic verification data in a public repository. “We can't fix the past, but we can make sure the next generation of papers is built on a solid foundation,” the editor-in-chief wrote in an accompanying editorial.

The National Institutes of Health followed suit. In its 2025 supplement guide for animal research, the agency added a section on substrain reporting, recommending that investigators verify the substrain of their animals at the start of a project and at six-month intervals. The guide also advises reviewers to flag any manuscript that uses C57BL/6 without specifying J or N.

“This is one of those rare moments where a single, concrete failure leads to systemic change,” says Eisen. “Usually, replication crises are diffuse—you can't point to one cause. Here, we have a clear culprit and a cheap fix.”

Practical Fixes That Cost Pennies but Save Careers

The fix is indeed cheap. Genotyping three SNP markers—rs13480097, rs13480100, and rs13480104—can distinguish J from N with near-perfect accuracy. The cost is roughly US$5–8 per sample, including DNA extraction. For a typical study using 50 mice, that is US$250–400—less than the price of a single animal from some vendors.

Since 2024, major mouse vendors have begun including substrain certificates of analysis with shipments. Charles River now labels every cage with the substrain and the date of the last genetic quality check. Jackson Laboratory has long done this, but the new policy ensures that distributors and end users receive the information automatically.

Lab protocols are also changing. Several large neuroscience labs have adopted a policy of running a substrain check every six months on a sample of their breeding colony. “It takes half a day and costs a few hundred dollars,” says Johnson. “If it saves even one retraction, it pays for itself a hundred times over.”

Preprint servers are getting involved too. bioRxiv and medRxiv now allow authors to tag their submissions with a “substrain” metadata field, making it searchable. “We want to make it easy for authors to do the right thing,” says Richard Sever, co-founder of bioRxiv. “If the metadata is there, it becomes part of the permanent record.”

The broader implications extend beyond mice. The same issue of hidden variation applies to cell lines, reagents, and software. For example, a 2023 study found that 30% of commonly used human cell lines are misidentified or contaminated. In neuroscience, different batches of antibodies can produce opposite results. The mouse substrain story is part of a larger pattern: the devil is in the details of provenance.

Lee's team is now developing a publicly accessible database that maps substrain usage in published studies. The database, funded by a small grant from the National Science Foundation, will allow researchers to check whether their own results might be confounded by substrain effects. “We want to make it easy for anyone to see if their field has a substrain problem,” Lee says. “The first step is awareness.”

The retraction also had personal consequences. One of the original paper's co-authors, a junior faculty member, had built a grant proposal around the findings. The retraction forced a rewrite. “It was devastating,” the researcher said in an interview, speaking on condition of anonymity. “But I'd rather know now than build a career on something that isn't real.”

The field is now grappling with the implications. Some journals have begun requesting substrain information for previously published papers. A few labs have started reanalyzing their old data using substrain as a covariate. “It's a lot of work,” Johnson admits. “But it's better than having to retract a paper five years from now.”

How many published papers might be affected? Lee's team estimates that roughly 200–400 studies using C57BL/6 mice without substrain verification could be vulnerable to undetected substrain effects. “That's a conservative number,” Lee says. “It only includes papers that explicitly state they used C57BL/6 without specifying J or N. The real number could be much higher.”

The broader lesson is that even the most mundane details of experimental design—the source of an animal, the lot number of a reagent, the version of a software library—can carry hidden variance. The replicability crisis in psychology and neuroscience has many causes, but some of them are surprisingly concrete. The mouse substrain story is a reminder that the path from preprint to settled claim is not just about statistical rigor or sample size. It is also about knowing exactly what you are working with, down to the last letter of its name.

As Lee puts it: “We spent millions of dollars and years of effort to learn something we could have known for five bucks and a PCR machine. The next time someone tells you that a variable is too small to matter, ask them to check.”