Everyday Scientist

The Case for Preregistration in Cell Biology: A Practical Guide

The basics of the scientific method—hypothesize, test, revise—rely on prediction. By forming a hypothesis only after seeing the data, we fail to severely test our ideas. I encourage cell biologists to preregister their study designs and statistical analyses. I also provide practical tutorials for making this a reality.

Preregistration will Bolster Cell Biology

Preregistration has helped add rigor to various fields, including psychology and clinical trials (Nosek et al., 2018). But in cell biology, molecular biology, and other basic research, preregistration has not become popular, at least partially due to a general belief that it would be overly burdensome and impractical in exploratory experiments. Here, I attempt to dispel those myths and provide practical steps for cell biologists to preregister their experiments and data analysis. Instead of hindering research activity, preregistration can aid experimentalists in designing better experiments, reducing waste, and adding rigor to their science.

While preregistration may sound daunting, it need not be. Most of us already write down a basic experimental plan in our lab notebook before we step into the lab, so uploading that to a repository (public or private) is a small effort. But that simple gesture could encourage researchers to flesh out their plans and more carefully consider potential pitfalls, like not having a sufficient sample size. There is nothing more annoying than completing weeks in the lab only to realize that you forgot an essential control or really needed an additional replicate to get a statistically significant result. Better planning can help avoid wasted effort, and preregistration helps. Getting feedback on your plan with labmates and colleagues is easier with a formal plan, and preregistration ensures that the PI and all the researchers on the project are on the same page.

We all know that arranging the data analysis procedure before seeing the results is good scientific hygiene. Preregistration forces us to make it a habit to plan the analysis, thereby reducing the temptation to p- and N-hack our way to “significance.” Furthermore, listing a priori plans and any deviations in the final paper adds transparency and trust to scientific publishing. Finally, committing to disclose results before the outcome is known helps alleviate the problem of selective reporting in scientific publishing.

Ancillary Benefits

Taking the time to plan an adequate study design before stepping into the lab is of course beneficial in itself. But preregistering an analysis plan could actually increase statistical power and even reduce the number of required samples. For instance, if the hypothesis predicts a directional effect of a treatment, researchers can plan for a one-sided t-test, which requires fewer samples to achieve the same power (Lakens, 2022). While it wouldn’t be fair to opt for a one-sided test after collecting the data and learning the direction of the effect, documenting the bold prediction of a directional effect beforehand is in some ways superior to the standard two-sided test. (Convincing reviewers of this may be a challenge, but they should look favorably on the transparency and rigor of preregistration.)

Preregistration can further reduce the burden on cell biologists by allowing them to “peek” at their data partway through completion of an experiment, providing the opportunity to stop data collection with fewer replicates if the effect is sufficiently large. “N-hacking” is when researchers arbitrarily add samples until they reach a “significant” result, but with proper planning using sequential analysis—correcting the α level for multiple peeks—one can possibly stop early while maintaining the desired false positive rate (Lakens, 2022; Schott et al., 2019). A major benefit of preregistering a sequential analysis is that the expected number of biological replicates is actually smaller than the maximum sample size you would plan, because peeking allows the possibility of stopping the experiment early with an already significant p-value (Lakens, 2022). This is a common approach in clinical trials, and cell biologists could benefit from borrowing this statistically rigorous way to make our lives easier. (Note that at very small sample sizes, like n=3 per stage per condition, corrections like Pocock have inflated false-positive rates. Rom-McTague (Rom and McTague, 2020) is a modified Pocock correction for small n.)

Finally, preregistering a plan can help make null results more credible to editors and reviewers, making it easier to publish all your results. Communicating the full picture, even the experiments that didn’t match your expectations, strengthens the integrity of your science by allowing you to tell the full story. Unfortunately, journals often make it (unofficially) difficult to publish results that do not have statistical significance or are otherwise ambiguous. Preregistration gives authors ammunition to justify including null results in their published findings.

Encouraging Preregistration

Doing robust science is a reward in itself, and the vast majority of researchers already strive to do science right. If preregistration became a norm—just like displaying scale bars on microscopy images, providing raw data, uploading structural data to PDB, sharing code on GitHub, reporting sample sizes, including loading controls for Westerns, and showing uncropped gels in the supporting information—then scientists would eventually preregister as a habit.  

But some external encouragement couldn’t hurt. Journals and reviewers could ask for preregistration documentation, and reward authors who can produce such plans. For example, when reviewer 2 asks for additional experiments or more replicates, editors should be emboldened by a preregistration to politely disregard such demands (when they are indeed unreasonable). Furthermore, editors should feel more willing to accept negative or ambiguous results when the experiment was preregistered. And journals should prominently display a link or badge on papers that provide preregistration documentation.

A Practical Guide to Preregistering

The OSF repository provides resources for recording preregistration reports (Open Science Framework, 2023), including what to do when plans change, and others have expanded on the benefits and practical application of preregistration in small studies (Johnson and Cook, 2019). Here, we provide a crude template (Box 1) for a simple preregistration. See Box 2 for example paragraphs describing the preregistration in the final manuscript. As your plans develop while you traverse a project, you can augment the preregistered plan with additional experiments. Because most cell biology papers include several smaller experiments, the preregistered plan can include descriptions of multiple experiments, or you may opt to submit a separate short plan for each experiment.

Common Objections

I don’t want my competitors knowing my experiments before I even start.

Preregistrations need not be public to be effective. Depositing a preregistration privately (or even simply writing it down, timestamping it, and storing it on a local computer) still keeps researchers accountable and the research design more robust against post-hoc statistical exploration. The plan can eventually be made public to peer reviewers and readers of the ultimate publication. That said, making a plan public—or at least sending it to some trusted colleagues—offers an opportunity to get helpful feedback on an experiment before you even enter the lab.

My experiments are too exploratory to plan ahead.

It probably doesn’t make much sense to preregister pilot studies and research that is still in the beginning, exploratory stages. So maybe 90% of the work you do in the lab—culturing and observing cells, developing assays, purifying proteins, making and characterizing reagents, performing chemical analyses, and running pilot experiments—doesn’t need a preregistration. But before you sit down to start an actual experiment, it is worth planning: write down, discuss, and preregister how you will tackle the project. It is challenging to recognize precisely when research shifts from exploratory to a confirmatory experiment, and I don’t have a perfect prescription for identifying which work deserves preregistration, but most scientists know in their gut when their efforts are destined for a figure in a paper. When in doubt, just fill out the preregistration template (Box 1); you can always discard it if it turns out it wasn’t needed or update it as necessary as the project progresses. That shouldn’t take too much time, and it might help focus your attention on deficits in your plan early on.

This is too much work.

Preregistering adds little administrative burden, since the hypothesis, research plan, and often even the statistical analyses have already been laid out in research grant applications, lab notebooks, or thesis committee presentations. But taking the simple steps of formally organizing and uploading these plans to a repository ensures you have a clear strategy and helps avoid common pitfalls of poor experimental planning. Each experiment can be preregistered on its own, so there’s no need to decide what the ultimate paper will be during the early planning stages.

This will lock me into a plan that might send me down the wrong path.

Preregistration is a plan, not a prison (Open Science Framework, 2023). If plans change before collecting the data, it’s as simple as updating the preregistered documents. If you perform unplanned tests or data analysis during or after the experiment, you can transparently report deviations from the plan and justifications in the final paper (Lakens, 2024). If you make an unexpected discovery or do additional experiments not listed in your original preregistration, you can either update the preregistration or simply note in the final manuscript that that particular experiment had not been listed in the original preregistered plan. This adds transparency and empowers the reader with useful information. “Having  preregistered  your  study  does  not  prevent  you from  publishing  any  analysis  you  believe  is  interesting  or informative. All that preregistration does is clarify what tests were formulated a priori and what tests were not” (Forstmeier et al., 2017).

What if the experiment doesn’t give the result I expected? Am I committed to trying to publish negative results?

Preregistration doesn’t obligate you to publish every result: you’re still free to tell the story the way you wish. The difference is that, when you document your full process, the readers will know that the findings are not cherry-picked, adding credibility and transparency to whatever you ultimately publish.

My Personal Experience with Preregistration

This is where I am seeking input from anyone who has ever preregistered their experiments before. For my next experiment, I will preregister the plan and the analysis approach and include that when we submit to a journal. But that will be months away. Has anyone already done this and want to expound upon the process and the outcome?

Conclusion

The value of preregistration has been demonstrated for years in clinical research and other fields, and it is time to apply this proven technique to cell biology. Of course, preregistration is not magic: it will not singlehandedly solve the replication crisis, eliminate fraud, or make all research findings true. Improving science will take a concerted effort by journals, authors, and funders to better incentivize transparent protocols and more rigorous testing of findings (Forstmeier et al., 2017; Kohrs et al., 2023; Lord et al., 2026). But this small yet powerful step would increase rigor in scientific research and help strengthen the literature. Ultimately, if you completely ditch your preregistered plan and don’t even mention it in your paper, while this is dishonest, it’s no worse off than the status quo. So there’s really no harm in experimenting with preregistration.

References

• Forstmeier, W., E.-J. Wagenmakers, and T.H. Parker. 2017. Detecting and avoiding likely false-positive findings – a practical guide. Biol. Rev. Camb. Philos. Soc. 92:1941–1968. doi:10.1111/brv.12315.
• Johnson, A.H., and B.G. Cook. 2019. Preregistration in Single-Case Design Research. Except. Child. 86:95–112. doi:10.1177/0014402919868529.
• Kohrs, F.E., S. Auer, A. Bannach-Brown, S. Fiedler, T.L. Haven, V. Heise, C. Holman, F. Azevedo, R. Bernard, A. Bleier, N. Bössel, B.P. Cahill, L.J. Castro, A. Ehrenhofer, K. Eichel, M. Frank, C. Frick, M. Friese, A. Gärtner, K. Gierend, and T.L. Weissgerber. 2023. Eleven strategies for making reproducible research and open science training the norm at research institutions. eLife. 12. doi:10.7554/eLife.89736.
• Lakens, D. 2022. Improving Your Statistical Inferences.
• Lakens, D. 2024. When and how to deviate from a preregistration. Collabra: Psychology. 10. doi:10.1525/collabra.117094.
• Lord, S.J., A. Charles-Orszag, K. Skruber, R.D. Mullins, and A. Rehfeld. 2026. Peer replication : A new tier of science built on reproducibility. EMBO Rep. doi:10.1038/s44319-026-00705-8.
• Nosek, B.A., C.R. Ebersole, A.C. DeHaven, and D.T. Mellor. 2018. The preregistration revolution. Proc Natl Acad Sci USA. 115:2600–2606. doi:10.1073/pnas.1708274114.
• Open Science Framework. 2023. Preregistration – OSF Support.
• Rom, D.M., and J.A. McTague. 2020. Exact critical values for group sequential designs with small sample sizes. J. Biopharm. Stat. 30:752–764. doi:10.1080/10543406.2020.1730878.
• Schott, E., M. Rhemtulla, and K. Byers-Heinlein. 2019. Should I test more babies? Solutions for transparent data peeking. Infant Behav. Dev. 54:166–176. doi:10.1016/j.infbeh.2018.09.010.

____________________________________________________

____________________________________________________

Box 2: Example paragraphs for eventual manuscript

We preregistered our research plans and statistical analyses (link to plans).

Based on a small pilot experiment, we predicted that abc drug treatment would reduce cell speeds relative to control, so we proposed a one-tailed t-test with α=0.05. Based on the Cohen’s effect size of d=2 from the pilot experiment, we proposed n=8 biological replicates (4 control and 4 treated), which should give a power of around 80%. After the experiment, the p-value was 0.043. However, we noticed after the data was collected that there was substantial variation in the speed of the control cells on different days. Because control and treated samples were split from the same flask of cells for each day’s experiment, we subsequently concluded that a paired t-test was the more appropriate approach and therefore departed from our preregistered analysis plan. The one-sided paired t-test resulted in p=0.011. Results are shown in Fig 2.

Also, after observing the full results, we noted that the nuclei seemed smaller in the drug-treated samples. Although not in our original preregistered plan, we decided to also compare the size of the nuclei in control vs. treated. We blinded the filenames and measured the area of the nuclei of all the cells in the dataset. The two-tailed paired t-test resulted in a p=0.034. Results are shown in Fig 3.

For the internalization experiment, we predicted that knockdown of gene xyz would abrogate endocytosis. We proposed a one-sided paired t-test with α=0.05 and planned n=6 biological replicates (3 control paired with 3 treated). However, our results showed no obvious trend in the internalization rate and the p-value was nonsignificant at p=0.31. Although the preregistered plan calculated that the experiment was adequately powered to detect a large biologically meaningful effect, the sample-to-sample variance was larger than we had anticipated, so we cannot conclude definitively that knockdown of gene xyz does not alter endocytosis rates, but it does set a floor for the effect size. Future tests of this hypothesis will need to either reduce the experimental noise or increase the number of replicates. This null result is reported in Fig S1.

For the neuron experiment, we proposed using an unpaired design, and we planned to use male mice from the same litter. We proposed to sacrifice control and knockout mice, fix and stain brain tissue, and image the average intensity of xyz marker in 100 neurons from each mouse. Sample collection was performed blinded with respect to the identity of the sample. We proposed a two-tailed t-test to calculate the p-value, because we did not know whether the marker would increase or decrease with gene knockout. To potentially reduce the number of mice needed, and based on an anticipated effect size of approximately 1.5, we planned a two-stage sequential analysis with n=5 mice per stage per condition; to account for multiple assessments of the data, we proposed the Rom-McTague α of 0.0285 for each stage, for a cumulative α of 0.05. After n=5 knockdown and 5 control mice, we calculated p=0.024, so we halted the experiment according to our plan. Results are shown in Fig 5.

PromoPlot

This is hilarious: https://arxiv.org/abs/2503.24254

end grant writing; award prizes for papers instead

My proposal for funding academic research science: instead of funding grants, just award financial prizes for good papers.

Implementation

Award half the prize upon publication (or preprint); award the second half upon an independent replication of the results, 25% to the original authors and 25% to the replicators. (If a good-faith replication attempt fails, the replicators would still receive their portion, but the original authors would not receive their replication bonus. The measure for the quality of the replication would be whether the work could be published in a reputable journal.)

Prize winning papers would be chosen by rotating committees of professors following a rubric announced prior to the prize committee meeting (replacing grant study sections). Different committees could be assembled for different research topics, with appropriately tailored rubrics. Any paper published in the prior 12 months would be eligible during each prize committee meeting (there could be multiple committees convened each year, giving any given paper multiple chances at winning). Some committees would be assembled to identify “sleeper” papers: works that are up to 5 years old that had been overlooked by previous committees but have proven to be especially valuable.

The prize amount could vary by field or even by paper. A typical R01 grant is $150k+ annually for 5 years. So prizes could vary from $200k up to $1M+ for a major tour-de-force paper with substantial impact. The goal would be to distribute the same pot of money to approximately the same number of labs in roughly the same proportion, but with a much smaller bureaucratic burden and with better incentives for reproducibility.

There would probably need to be some cap on the total prize money an individual PI could receive at any one time. If they are over the cap, then either their papers would simply not be eligible for a prize that round or their prize money would flow to the collaborators on that paper or department.

Prize money would still go to the PI and lab via the university, not as a blank check to an individual professor. Indirects/overhead would still be applied. The goal would be to have approximately the same amount of money flowing to the university and the labs, but with a less onerous process.

Advantages

The prize approach accomplishes the following:

• Rewards actual outputs rather than promises.
• Creates strong incentives for reproducible work.
• Incentivizes impactful, innovative results that are convincing and accepted by the scientific community.
• Frees up professors’ time to focus on paper writing and mentoring instead of grant writing.
• Reduces the workload for university grant offices, which currently spend considerable time reviewing and approving proposals before submission, thereby reducing costs and freeing administrative staff at universities and funding agencies to focus on other support functions.
• Removes back-and-forth about budgets, justifications, biosketches, data-sharing policies, etc.
• Does away with byzantine formatting requirements.
• Obviates¹ the need for progress reports and renewals throughout the grant period.

Professors spend a huge amount of time writing grants, and university grant offices dedicate substantial resources helping researchers navigate complex formatting and compliance requirements, including page lengths, fonts, hyperlinks, sections, etc. Academic researchers currently focus on publishing their work, so grants should reinforce that drive, not distract PIs with hundreds of hours of grant writing. Prize committees would replace grant review panels (roughly equivalent work), while reducing the administrative burden on both university grant offices and granting agencies.

The current grant funding structure purports to be forward-looking, providing funding for good ideas to be implemented. But actual grants require not only a good track record for the PI, but also a large amount of preliminary data. So, in effect, grants fund a lot of work that’s already been completed, and the awarded money usually gets spent on lab work outside the scope of the original proposal. So, while appearing to be prospective, grants are often retrospective in practice. Granting prizes for good papers simply makes official how funding already flows in reality.

How to fund research before winning prizes

The central challenge would be providing runway funding, especially for early-career researchers who have no eligible papers yet. With fewer resources needed for proposal preparation support, some overhead funds could be redirected to departmental slush funds. Departments could then provide startup packages, bridge funding between prizes, or support for labs that demonstrate progress toward publishable work.

This would shift evaluation of proposed research from granting agencies to local departments, where a lab’s potential is more knowable. However, departments are often rife with unhealthy politics and lack sufficient guardrails against favoritism and bias. Whether departmental evaluation would be less burdensome than current grants – and whether freed-up overhead would provide sufficient funding – remains an open question.

Alternatively, granting agencies could dedicate some funds for prizes awarded specifically to a professor’s first paper, or maintain a smaller pot of traditional proposal-style funding for early-career researchers. But as it stands, professors who don’t publish within their first few years are unlikely to receive traditional grants or be awarded tenure anyway, so the prize approach doesn’t necessarily create more pressure than the current system.

Downsides

Not every good paper would win a prize. Yes, of course there would still be disappointing results: inevitably, some deserving papers would fail to attract money because the subject matter is out of fashion or reviewer bias or a host of other unfair reasons. But with a limited amount of money and an effectively unlimited number of professors, it’s impossible to fund everyone, so someone will always be neglected. Winning a prize would be no more unpredictable than receiving a good score on a grant. My proposal is no less fair than the whims of the current granting process.

Incentivizes short-term work. Scientists would be less motivated to tackle a longer-term project, even if it is promising. The department or other funding agency would need to step in to float the project for years until it comes to fruition. And if whoever funds that float requires proposals to have the same level of detail and complexity as current grant proposals, we’d be back to square one. But proposals for future work need not be as onerous as current NIH regulations, especially if funded by the department or by private organizations like HHMI. Another source of funding for longer-term projects may actually come from paper prizes: professors who have won multiple prizes in the recent past would have ample money to fund collaborations and longer projects.

Disincentivizes high-risk research. Researchers might be wary of taking on high-risk/high-reward projects, because, if the experiments fail, they couldn’t expect a prize. That said, if such a project succeeds, then the prize money would likely be larger. Such is the nature of high-risk/high-reward. Therefore, the prize money for impactful work would need to be sufficiently large to encourage at least a portion of researchers to take on projects that are higher risk. Furthermore, if researchers are able to publish details of a failed “moonshot” that advanced the field despite not succeeding, such a paper would be eligible for a prize. The flexibility of prize funding might actually enable more ambitious research, since successful researchers would have discretion over how to deploy their winnings without the constraints of grant-specified aims.

Disincentivizes collaborations. Collaborators would likely negotiate some cut of a potential prize ahead of time, or there would be a standard formula (e.g. 75% to the last author and the remainder of the prize to the other PI collaborators on the paper). Maybe PIs would be less willing to add collaborators to a project, for fear that they may need to share some of the prize money with other PIs. Maybe there could be a bonus to help support collaborators.

Expensive projects. This would not work for huge projects like clinical trials or massive particle physics consortiums, which would need ample funding before the experiments or construction actually begin. But those could continue to be funded in the current fashion.

What will professors do with their freed up time? Part of the job of a professor is to write grants. If they suddenly don’t need to write grants, will they actually spend more time teaching and mentoring, or will they just work less? Not sure. That’s a valid point.

Moving forward

The next step would be to run a pilot. Initially, 20% of NIH and NSF funding could be converted to this model in select fields that would be amenable, then evaluated after 5 years. At that time, the percentage of funding provided via the prize mechanism might be increased or decreased based on the outcome of the pilot. It’s unlikely that granting agencies would ever completely eliminate the traditional grant proposal, but the health of academic science could be strengthened if even a minority of the funding rewarded impactful, reproducible papers.

___________________

1. Thesaurus, don’t fail me now. ↩︎

eLife’s new publishing policy

Let me preface this post with the admission that I’m likely wrong about my concerns about eLife’s new model. I was actually opposed to preprints when I first heard about them in 2006!

The Journal eLife and Mike Eisen announced it’s new model for publishing papers:

• Authors post preprint.
• Authors submit preprint to eLife.
• eLife editorial board decides whether to reject the manuscript or send out for review.
• After reviews, the paper will be published no matter what the reviews say. The reviews and an eLife Assessment will be published alongside the paper. At this point, the paper has a DOI and is citable.
• Authors then have a choice of whether to revise their manuscript or just publish as-is.
• When the authors decide the paper is finalized, that will become the “Version of Record” and the paper will be indexed on Pubmed.

Very interesting and bold move. The goal is to make eLife and its peer review not a gatekeeper of truth, but instead a system of evaluating and summarizing papers. Mike Eisen hopes that readers will see “eLife” and no longer think “that’s probably good science” and instead think “oh, I should read the reviews to see if that’s a good paper.”

Potential problems

But here are my primary concerns

1. This puts even more power in the hands of the editors to effectively accept/reject papers. And this process is currently opaque, bias-laden, and authors have no recourse when editors make bad decisions.
2. The idea that the eLife label will no longer have prestige is naive. The journal has built a strong reputation as a great alternative to the glam journal (Science, Nature, Cell) and that’s not going away. For example, people brag when their papers are reviewed in F1000, and I think the same will apply to eLife Assessments: readers will automatically assume that a paper published in eLife is high-impact, regardless of what the Assessment says.
3. The value that eLife is adding to the process is diminishing, and the price tag is steep ($2000).
4. The primary problem I have with peer review is that it is simultaneously overly burdensome and not sufficiently rigorous. This model doesn’t substantially reduce the burden on authors to jump through hoops held by the reviewers (or risk a bad eLife Assessment). It also is less rigorous by lowering the bar to “publication.”

Solutions

Concern #1: I think it’s a step in the wrong direction to grant editors even more power. Over the years, editors haven’t exactly proven themselves to be ideal gatekeepers. How can we ensure that the editors will act fairly and don’t get attracted by shiny objects? That said, this policy might actually put more of a spotlight on the desk-rejection step and yield change. eLife could address this concern in various ways:

• The selection process could be a lottery (granted, this isn’t ideal because finding editors and reviewers for a crappy preprint will be hard).
• Editors could be required to apply a checklist or algorithmic selection process.
• The editorial process could be made transparent by publishing the desk rejection/acceptace along with the reasons.

Concern #2 might resolve itself with time. Dunno. Hard to predict how sentiment will change. But I do worry that eLife is trying to change the entire system, while failing to modify any of the perverse incentives that drive the problems in the first place. But maybe it’s better to try something than to do nothing.

Concern #3 is real, but I’m sure that Mike Eisen would love it if a bunch of other journals adopted this model as well and introduced competition. And honestly, collating and publishing all the reviews and writing a summary assessment of the paper is more than what most journals do now.

Journals should be better gatekeepers

But #4 is pretty serious. The peer review process has always had to balance being sufficiently rigorous to avoid publishing junk science with the need to disseminate new information on a reasonable timescale. Now that preprinting is widely accepted and distributing results immediately is super easy, I am less concerned with latter. I believe that the new role of journals should be as more exacting gatekeepers. But it feels like eLife’s policy was crafted exclusively by editors and authors to give themselves more control, reduce the burden for authors, and shirk the responsibility of producing and vetting good science.

There are simply too many low-quality papers. The general public, naive to the vagaries of scientific publishing, often take “peer-reviewed” papers as being true, which is partially why we have a booming supplement industry. Most published research findings are false. Most papers cannot be replicated. Far too many papers rely on pseudoreplication to get low p-values or fail to show multiple biological replicates. And when was the last time you read a paper where the authors blinded their data acquisition or analysis?

For these reasons, I think that the role of a journal in the age of preprints is to better weed out low-quality science. At minimum, editors and peer reviewers should ensure that authors followed the 3Rs (randomize, reduce bias, repeat) before publishing. And there should be a rigorous checklist to ensure that the basics of the scientific process were followed.

Personally, I think the greatest “value-add” that journals could offer would be to arrange a convincing replication of the findings before publishing (peer replication), then just do away with the annoying peer review dog-and-pony show altogether.

Conclusion

We’ll have to wait and see how this new model plays out, and how eLife corrects stumbling blocks along the way. I have hope that, with good editorial team and good practices/rules around the selection process, eLife might be able to pull this off. Not sure if it’s a model that will scale to other, less trustworthy journals.

But just because this isn’t my personal favorite solution to the problem of scientific publishing, that doesn’t mean that eLife’s efforts won’t help make a better world. I changed my mind about the value of preprints, and I’ll be happy to change my mind about eLife’s new publishing model if it turns out to be a net good!

Replace Peer Review with “Peer Replication”

UPDATE: EMBO Reports article here and a write-up in Nature here.

As I’ve posted before and many others have noted, there is a serious problem with lack of adequate replication in many fields of science. The current peer review process is a dreadful combination of being both very fallible and also a huge hurdle to communicating important science.

Instead of waiting for a few experts in the field to read and apply their stamp of approval to a manuscript, the real test of a paper should be the ability to reproduce its findings in the real world. (As Andy York has pointed out, the best test of a new method is not a peer reviewing your paper, but a peer actually using your technique.) But almost no published papers are subsequently replicated by independent labs, because there is very little incentive for anyone to spend time and resources testing an already published finding. That is precisely the opposite of how science should ideally operate.

Let’s Replace Traditional Peer Review with “Peer Replication”

Instead of sending out a manuscript to anonymous referees to read and review, preprints should be sent to other labs to actually replicate the findings. Once the key findings are replicated, the manuscript would be accepted and published.

(Of course, as many of us do with preprints, authors can solicit comments from colleagues and revise a manuscript based on that feedback. The difference is that editors would neither seek such feedback nor require revisions.)

Along with the original data, the results of the attempted replication would be presented, for example as a table that includes which reagents/techniques were identical. The more parameters that are different between the original experiment and the replication, the more robust the ultimate finding if the referees get similar results.

Incentives

What incentive would any professor have to volunteer their time (or their trainees’ time) to try to reproduce someone else’s experiment? Simple: credit. Traditional peer review requires a lot of time and effort to do well, but with zero reward except a warm fuzzy feeling (if that). For papers published after peer replication, the names of researchers who undertook the replication work will be included in the published paper (on a separate line). Unlike peer review, the referees will actually receive compensation for their work in the form of citations and another paper to include on their CV.

Why would authors be willing to have their precious findings put through the wringer of real-world replication? First and foremost, because most scientists value finding truth, and would love to show that their findings hold up even after rigorous testing. Secondly, the process should actually be more rewarding than traditional peer review, which puts a huge burden on the authors to perform additional experiments and defend their work against armchair reviewers. Peer replication turns the process on its head: the referees would do the work of defending the manuscript’s findings.

Feasible Experiments

There are serious impediments to actually reproducing a lot of findings that use seriously advanced scientific techniques or require long times or a lot of resources (e.g. mouse work). It will be the job of editors—in collaboration with the authors and referees—to determine the set of experiments that will be undertaken, balancing rigor and feasibility. Of course, this might leave some of the most complex experiments unreplicated, but then it would be up to the readers to decide for themselves how to judge the paper as a whole.

What if all the experiments in the paper are too complicated to replicate? Then you can submit to JOOT.

Ancillary Benefits

Peer replication transforms the adversarial process of peer review into a cooperation among colleagues to seek the truth. Another set of eyes and brains on an experiment could introduce additional controls or alternative experimental approaches that would bolster the original finding.

This approach also encourages sharing experimental procedures among labs in a manner that can foster future collaborations, inspire novel approaches, and train students and postdocs in a wider range of techniques. Too often, valuable hands-on knowledge is sequestered in individual labs; peer replication would offer an avenue to disseminate those skills.

Peer replication would reduce fraud. Often, the other authors on an ultimately retracted paper only later discover that their coworker fabricated data. It would be nearly impossible for a researcher to pass off fabricated data or manipulated images as real if other researchers actually attempt to reproduce the experimental results. 

Potential Problems

One serious problem with peer replication is the additional time it may take between submission and ultimate publication. On the other hand, it often takes many months to go through the traditional peer review process, and replicating experiments may not actually add any time in many cases. Still this could be mitigated by authors submitting segments of stories as they go. Instead of waiting until the entire manuscript is polished, authors or editors could start arranging replications while the manuscript is still in preparation. Ideally, there would even be a  journal-blind mechanism (like ReviewCommons) to arrange reproducing these piecewise findings.

Another problem is what to do when the replications fail. There would still need to be a judgement call as to whether the failed replication is essential to the manuscript and/or if the attempt at replication was adequately undertaken. Going a second round at attempting a replication may be warranted, but editors would have to be wary of just repeating until something works and then stopping. Pre-registering the replication plan could help with that. Also, including details of the failed replications in the published paper would be a must.

Finally, there would still be the problem of authors “shopping” their manuscript. If the replications fail and the manuscript is rejected, the authors could simply submit to another journal. I think the rejected papers would need to be archived in some fashion to maintain transparency and accountability. This would also allow some mechanism for the peer replicators to get credit for their efforts.

Summary of Roles:

• Editor:
  • Screen submissions and reject manuscripts with obviously flawed science, experiments not worth replicating, essential controls missing, or seriously boring results.
  • Find appropriate referees.
  • With authors and referees, collaboratively decide which experiments the referees should attempt to replicate and how.
  • Ultimately conclude, in consultation with referees, whether the findings in the papers are sufficiently reproducible to warrant full publication.
• Authors:
  • Write the manuscript, seek feedback (e.g. via bioRxiv), and make revisions before submitting to the journal.
  • Assist referees with experimental design, reagents, and even access to personnel or specialized equipment if necessary.
• Referees:
  • Faithfully attempt to reproduce the experimental results core to the manuscript.
  • Optional: Perform any necessary additional experiments or controls to close any substantial flaws in the work.
  • Collate results.
• Readers:
  • Read the published paper and decide for themselves if the evidence supports the claims, with the confidence that the key experiments have been independently replicated by another lab.
  • Cite reproducible science.

How to Get Started

While it would be great if a journal like eLife simply piloted a peer replication pathway, I don’t think we can wait for Big Publication to initiate the shift away from traditional peer review. Maybe the quickest route would be for an organization like Review Commons to organize a trial of this new approach. They could identify some good candidates from bioRxiv and, with the authors, recruit referees to undertake the replications. Then the entire package could be shopped to journals.

I suspect that once scientists see peer replication in print, it will be hard to take seriously papers vetted only by peer review. Better science will outcompete unreproduced findings.

(Thanks Arthur Charles-Orszag for the fruitful discussions!)

avoiding bias by blinding

Key components of the scientific process are: controls, avoiding bias, and replication. Most scientists are great at controls, but without the other two, we’re simply not doing science.

The lack independent samples (and thus improper inflation of n) and the failure to blind experiments are too common. The implications of these mistakes, especially when combined in one study, mean that many published cell biology results are likely artifact. Generating large datasets, even with a slight bias, can quickly yield “significant” results out of noise. For example, see the “NHST is unsuitable for large datasets” section of:

Szucs D, Ioannidis JPA. When Null Hypothesis Significance Testing Is Unsuitable for Research: A Reassessment. Front Hum Neurosci. 2017;11:390. https://pubmed.ncbi.nlm.nih.gov/28824397/

Now combine these common false positives with the inclination to publish flashy results, and we’ve made a recipe for unreliable scientific literature.

I do not condemn authors for these problems. Most of us have made one or all of these mistakes. I have. And I probably will again in the future. Science is hard. There is no shame in making honest mistakes. But we can all strive to be better (see my last section).

Failing to perform the data collection blinded

Blinding is just basic scientific rigor, and skipping this should be considered almost as bad as skipping controls.

Blinding samples during data collection and analysis is ideal. For data collection, it is usually as simple as having a labmate put tape over labels on your samples and label them with a dummy index. Insist that your coworker writes down the key, so later you can decode the dummy index back to the true sample information.

In cases where true blinding is impractical, the selection of cells to image/collect should be randomized (e.g. set random coordinates for the microscope stage) or otherwise designed to avoid bias (e.g. selecting cells using transmitted light if fluorescence the readout).

Failing to perform the data analysis blinded

Blinding during data analysis is generally very practical, even when the original data was not collected in a bias-free fashion. Ideally, image analysis would be done entirely by algorithms and computers, but often the most practical and effective approach is old-fashioned human eye. Ensuring your manual analysis isn’t biased is usually as simple as scrambling the image filenames.

I stumbled upon these ImageJ macros for randomizing and derandomizing image filenames, written by Martin Höhne: http://imagej.1557.x6.nabble.com/Macro-for-Blind-Analyses-td3687632.html

More recently, Christophe Leterrier directed me to Steve Royle‘s macro, which works very well: https://github.com/quantixed/imagej-macros#blind-analysis

There are probably some excellent solution using Python. Regardless of the approach you take, I would highly recommend copying all the data to a new folder before you perform any filename changes. Then test the program forward and backwards to confirm everything works as expected. Maybe perform analysis in batches, so in case something goes awry, you don’t lose all that work.

My “blind and dumb” experiment

There are many stories about unintended bias leading to false conclusions. Here’s mine: I was testing to see whether a drug treatment inhibited cells from crawling through a porous barrier by counting the number of cells that made it through the barrier to an adjacent well.

My partner in crime had labeled the samples with dummy indices, so I didn’t know which wells were treated and which were control. But I immediately could tell that there were more cells in one set of wells, so I presumed those were the control set. Fortunately, I had taken the extra precaution of randomizing the stage positions, so I didn’t let my bias alter the data collection. We then blinded the analysis by relabeling the microscopy images. I manually counted all the cells in each image.

We then unblinded the samples. At first, we were disappointed that the wells I had assumed were control turned out to be treated. Then we looked at the results. SURPRISE! My snap judgement at the beginning of the experiment had been precisely backwards: the wells I thought looked like they had sparser cells actually had significantly more on average. So it turned out that the drug treatment had indeed worked. Thankfully, I didn’t rely on my snap judgement nor allow that bias to influence the results.

Treating each cell as an n in the statistical analysis

This error plagues the majority of cell biology papers. Go scan a recent issue of your favorite journal and count the number of papers that have minuscule P values; invariably, the authors aggregated all the cell measurements from multiple experiments and calculated the t-test or ANOVA based on those dozens or hundreds of measurements. This is a fatal error.

It is patently absurd to consider neighboring cells in the same dish all treated simultaneously with the same drug as independent tests of a hypothesis.

If your neighbor told you that he ate a banana peal and it reversed his balding, you might be a little skeptical. If he further explained that he measured 1000 of his hair follicles before and after eating a banana peel and measured a P < 0.05 difference in growth rate, would you be convinced? Maybe it was just a fluke or noise that his hairs started growing faster. You would want him to repeat the experiment a few times (maybe even with different people) before you started believing.

Similarly, there are many reasons two dishes of cells might be different. To start believing that a treatment is truly effective, we all understand that we should repeat the experiment a few times and get similar results. Counting each cell measurement as the sample size n all but guarantees a small—but meaningless—P value.

Observe how dramatically different scenarios (on the right) yield the same plot and P value when you assume each cell is a separate n (on the left):

Elegant solutions include “hierarchical” or “nested” or “mixed effect” statistics. A simple approach is to separately pool the cell-level data from each experiment, then compare experiment-level means (n the case above, the n for each condition would be 3, not 300). For more details, please read my previous blog post or our paper:

Lord SJ, Velle KB, Mullins RD, Fritz-Laylin LK. SuperPlots: Communicating reproducibility and variability in cell biology. J Cell Biol. 2020;219(6):e202001064.
https://pubmed.ncbi.nlm.nih.gov/32346721

How do we fix this?

Professors need to teach their trainees the very basics about how to design experiments (see Stan Lazic’s book: Experimental Design for Laboratory Biologists) and perform analysis (see Mike Whitlock and Dolph Schluter’s book: The Analysis of Biological Data). PIs need to provide researchers with the tools to blind their experiments or otherwise remove bias. They need to ask for multiple biological replicates and correctly calculated P values. This does not require advanced understanding of statistics, just the basic understanding of the importance of repeating an experiment multiple times to ensure an observation is real.

Editors and referees need to demand correct data analysis. While asking researchers to redo an experiment isn’t really acceptable, requiring a reanalysis of the data after blinding or recalculating P values based on biological replicates seems fair. Editors should not even send manuscripts to referees if the above errors are not corrected or at least addressed in some fashion. Editors can offer the simple solutions listed above.

UPDATE: Also read my proposal to replace peer review with peer “replication.”

unbelievably small P values?

Check out our newest preprint at arXiv:

UPDATE: Now published in JCB: https://doi.org/10.1083/jcb.202001064

I’ve noticed a promising trend away from bar graphs in the cell biology literature. That’s great, because reporting simply the average and SD or SEM or an entire dataset conceals a lot of information. So it’s nice to see column scatter, beeswarm, violin, and other plots that show the distribution of the data.

But a concerning outcome of this trend is that, when authors decide to plot every measurement or every cell as a separate datapoint, it seems to trick people into thinking that each cell is an independent sample. Clearly, two cells in the same flask treated with a drug are not independent tests of whether the drug works: there are many reasons the cells in that particular flask might be different from those in other flasks. To really test a hypothesis that the drug influences the cells, one must repeat the drug treatment multiple times and check if the observed effect happens repeatably.

I scanned the latest issues of popular cell biology journals and found that over half the papers counted each cell as a separate N and calculated P values and SEM using that inflated count.

But that’s not to say the the variability cell-to-cell is unimportant! The fact that some cells in a flask react dramatically to a treatment and others carry on just fine might have very important implications in an actual body.

So we proposed “SuperPlots,” which superimpose sample-to-sample summary data on top of the cell-level distribution. This is a simple way to convey both variability of the underlying data and the repeatability of the experiment. It doesn’t really require any complicated plotting or programming skills. On the simplest level, you can simply paste two (or more!) plots in Illustrator and overlay them. Play around with colors and transparency to make it visually appealing, and you’re done! (We also give a tutorial on how we made the plots above in Graphpad Prism.)

Let me know what you think!

UPDATE: We simplified the figure:

Figure 1. Drastically different experimental outcomes can result in the same plots and statistics unless experiment-to-experiment variability is considered. (A) Problematic plots treat N as the number of cells, resulting in tiny error bars and P values. These plots also conceal any systematic run- to-run error, mixing it with cell-to-cell variability. To illustrate this, we simulated three different scenarios that all have identical underlying cell-level values but are clustered differently by experiment: (B) shows highly repeatable, unclustered data, (C) shows day-to-day variability, but a consistent trend in each experiment, and (D) is dominated by one random run. Note that the plots in (A) that treat each cell as its own N fail to distinguish the three scenarios, claiming a significant difference after drug treatment, even when the experiments are not actually repeatable. To correct that, “SuperPlots” superimpose summary statistics from biological replicates consisting of independent experiments on top of data from all cells, and P values were calculated using an N of three, not 300. In this case, the cell-level values were separately pooled for each biological replicate and the mean calculated for each pool; those three means were then used to calculate the average (horizontal bar), standard error of the mean (error bars), and P value. While the dot plots in the “OK” column ensure that the P values are calculated correctly, they still fail to convey the experiment-to-experiment differences. In the SuperPlots, each biological replicate is color-coded: the averages from one experimental run are yellow dots, another independent experiment is represented by gray triangles, and a third experiment is shown as blue squares. This helps convey whether the trend is observed within each experimental run, as well as for the dataset as a whole. The beeswarm SuperPlots in the rightmost column represent each cell with a dot that is color coded according to the biological replicate it came from. The P values represent an unpaired two-tailed t-test (A) and a paired two-tailed t-test for (B-D). For tutorials on making SuperPlots in Prism, R, Python, and Excel, see the supporting information.

electrically tunable lenses for microscopy

Electrically tunable lenses (ETLs) are polymeric or fluid-filled lenses that have a focal length that changes with an applied current. They have shown some great potential for microscopy, especially in fast, simple z-sweeps.

The above figure shows the ~120 um range of focal depths an ETL installed between the camera and a 40x objective (from reference 1). Note that this arrangement has the drawback of changing the effective magnification at different focal depths; however, this effect is fairly small (20%) and linear over the full range. For high-resolution z-stack imaging of cells, this mag change would not be ideal. But it should be correctable for imaging less sensitive to magnification changes. Basic ETLs cost only a few hundred dollars, a lot cheaper than a piezo stage or objective focuser. Optotune has a lot of information about how to add an ETL to a microscope.

Another cool application of an ETL is in light-sheet microscopy. A recent paper from Enrico Gratton (reference 2) used an ETL to sweep the narrow waist of a light sheet across the sample, and synchronize its motion to match the rolling shutter of a CMOS camera.

The main goal was to cheaply and simply create a light sheet that had a uniform (and minimal) thickness across the entire field of view. Previous low-tech methods to achieve this was to close down an iris, thus reducing the difference in thickness across the sample, but it also reduces the minimal waist size. The high-tech way to do this is creating “propagation-invariant” Bessel or Airy beams. These do not spread out as they propagate, like Gaussian beams do, but creating them and aligning them in microscopes is significantly more challenging.

Gratton’s cheap trick means one can create a flat and thin light sheet for the cost of an ETL and the complexity of synchronizing a voltage ramp signal to the CMOS rolling shutter readout. To be honest, I don’t 100% know how complicated or robust that is in practice. I’m just guessing that it’s simpler than a Bessel beam.

1. Wang, Z., Lei, M., Yao, B., Cai, Y., Liang, Y., Yang, Y., … Xiong, D. (2015). Compact multi-band fluorescent microscope with an electrically tunable lens for autofocusing. Biomedical Optics Express, 6(11), 4353. doi:10.1364/BOE.6.004353

2. Hedde, P. N., & Gratton, E. (2016). Selective plane illumination microscopy with a light sheet of uniform thickness formed by an electrically tunable lens. Microscopy Research and Technique, 00(April). doi:10.1002/jemt.22707

experimenting with preprints

We recently uploaded a preprint to bioRxiv. The goal was to hopefully get some constructive feedback to improve the manuscript. So far, it got some tweets and even an email from a journal editor, but no comments or constructive feedback.

I notice that very few preprints on bioRxiv have any comments at all. Of course, scientists may be emailing each other privately about papers on bioRxiv, and that would be great. But I think a open process would be valuable. F1000Research, for example, has a totally open review process, posting the referee reports right with the article. I might be interested in trying that journal someday.

UPDATE: In the end, we did receive a couple emails from students who had read the preprint for their journal club. They provided some nice feedback. Super nice! We also received some helpful feedback on another preprint, and we updated the manuscript before submitting to a journal. Preprints can be super useful for pre-peer review.

speck of dust

The scope room dustiness post reminded me of the hilarious story of the first report of second harmonic generation (SHG) of a laser. The authors presented a photographic plate that showed the exposure from the main laser beam, as well as a “small but dense” spot from the doubled beam,

See the spot? You won’t. Because the editor removed the spot, thinking it was a speck of dust on the plate. Ha!

When I first heard this story, I didn’t believe it. I assumed it was a contrast issue when the paper was scanned into a PDF. So I went to the library and found the original print version. No spot there, either!

That really made my day.

am i finished using Papers?

I’ve been using Papers for years. When Papers2 came out, I was quick (too quick) to jump in and start using it. It’s worst bugs got ironed out within a couple months, and I used it happily for a while. Papers2 would let you sync PDFs to your iPad for offline reading, but it was slow and a little clunky. Papers3 library syncing is not for offline reading and it is VERY slow and VERY clunky. And it relies on Dropbox for storage. The plus of this is that storage is free (as long as you have space in Dropbox); the downside is that they syncing isn’t clean and often fails.

Mendeley has proven itself the best at syncing your library and actual PDFs to the cloud (you have to pre-download individual files for offline reading you can sync all PDFs in iOS in settings). Papers PDF viewer is still better, but it’s not worth the hassle: Mendeley syncs cleanly and the reader is fine. Not only that, but Mendeley has sharing options that make managing citations possible when writing a manuscript with co-authors (as long as they’ll use Mendeley).

Mendeley is also better than Papers at automatically finding the metadata for the paper (authors, title, abstract, etc.). The program simply works (most of the time), so I’ve given up and finally started using it. Almost exclusively.

PubChase syncs with Mendeley and recommends related papers weekly. (Update: the recommendations update daily, and they send out a weekly email with updates from that week.) They also have some pretty nice features, like a beautiful viewer for some journals and alerts when papers in your library are retracted.

Readcube still has the best recommendations. And they update daily, unlike PubChase’s weekly. And you can tell which recommendations you’ve marked as read, so it’s very quick to scan the list. But that’s really where Readcube’s benefits end. The enhanced PDF viewing feature is nice (it shows all the reference in the sidebar), but not really worth the slow-down in scrolling performance. The program is just clunky still. (I thought Adobe was slow!) And there’s no iOS/Android app yet. It’s on its way, allegedly, but I need it now! Readcube is really taking off, so maybe in a year it will be perfect. But not yet.

Edit: Readcube has a new version of their desktop application. Maybe it’s faster? Wait, did the references sidebar disappear? No, wait, it’s there. Just not on by default.

readcube and deepdyve update

I just wanted to reiterate how great the ReadCube recommendations are. I imported all my PDFs and now check the recommendations every day. I often find great papers (and then later find them popping up in my RSS feeds).

Also, I wanted to let folks know that DeepDyve, the article rental site, is now allowing free 5-min rental of journal articles. Try it out!

PubReader review

I’ve reviewed several PDF reader/organizers, like ReadCube, Papers, and Mendeley. Currently, I use Papers for organizing my PDF library on my computer. I also like Papers a lot for reading PDFs, because it displays in full screen so well. But I’ve started using Mendeley for adding citations to Word documents, because it makes it really easy to collaborate with other people who have Mendeley.

Now check out PubReader! It’s really cool. Pubmed has the advantage that it requires all research publications resulting from NIH funding to be uploaded to their depository. And they don’t just grab a PDF; they get the raw text and figures and they format it their own way. I used to think that was silly and overkill, but now I see that that approach was genius: it now allows Pubmed to reformat the papers into more readable shapes and sizes … and they can reformat in the future when the old format becomes antiquated. You can’t really do that with a PDF.

It’s always been nearly impossible to read PDFs on a phone or an e-ink tablet like the basic Kindle. Now, with PubReader and the beta option to download the article in an ePub format (for reading in iBooks or Kindle or something), that option is here. Or on its way, at least.

PubReader on a computer:

PubReader on iPad:

ePub in iBooks:

Now PubReader just needs to display the references in an elegant way like ReadCube, and it will be the best!

It makes me think the future of reading and storing scientific papers is not the hard drive, but simply reading on online depositories. Pubmed allows you to create collection and star favorites, so you can just use Pubmed to store your collection of papers and never have to download a PDF again in your life!

readcube review

I recently tried Readcube, which is a PDF reader and organizer. I did so because Nature has been using it built into their site, and I like how it displaying PDFs. The article data downloads seamlessly for most papers, and  interface is quite beautiful:

The really cool feature is that Readcube automatically downloads the references and the supporting information documents and can display them at a click of a button. More importantly, it displays the references in the sidebar. It makes an excellent reading experience!

The final interesting feature is that Readcube offers recommendations based on your library. From my quick scan, the recommendations seem pretty good.

Other than that, Readcube is quite feature poor. It doesn’t have a way to insert citations into a Word document, like Papers and Mendeley does, although you can export to Endnote. I don’t see a way to read in full screen nor does it let you view two pages simultaneously, like Papers does.

The screenshot above is from Papers fullscreen view, which is how I really like to read PDFs.

But Readcube is still in beta, and they’re starting from a really nice starting point. I’m not ready to give up on Papers for reading (and I’ve been using Mendeley for Word citations, because it has really nice collaborative features). But I might try Readcube some more, mainly because of the awesome ability to see all the references and the paper simultaneously. I really wish I could mash Papers, Mendeley, and Readcube all together into one feature-rich program…

ActiveView PDF

Does anyone else love ACS’s ActiveView PDF viewer for reading PDFs and seeing reference? And Nature’s ReadCube, too. Great stuff.

Of course, after I scan the ActiveView, I still download the old-fashioned PDF and use Papers (or Mendeley) to read and manage my library.