[Version anglaise de l’article « Comment améliorer la reproductibilité scientifique ? »]
It is undeniable: science is facing a reproducibility crisis. It sounds simple, but the problem has numerous and multifaceted causes. Among them, errors made in good faith and fraud, often caused by high-pressure work environments. Another major cause is a persistent lack of available and transparent methods, codes, and raw data to facilitate study replication. An absence all the more surprising considering, according to science-sociologist Robert Merton, open access to research results is a very pillar of the scientific ethos– Merton refers to it as “communalism.” Scientific research needs to be accessible in order to be further interrogated, extended, and sometimes disproven by our peers. Without the ability to check and correct each other’s work, errors persist in the literature, ultimately undermining both scientific progress and public trust.
So, what can we do to improve the situation? Here are two (non-exhaustive) ideas!
Approach #1: Create opportunities to replicate published articles and recognize the effort involved
Several initiatives are organizing structured replication efforts at varying scales. Currently, a little over 3,000 of them are listed in the FORRT Library of Replication Attempts (FLoRA)! Before you get lost in the database, I have a few examples.
Perhaps one of the best-known efforts is the Reproducibility Project: Cancer Biology, which attempted to replicate 50 experiments from 23 high-impact articles in, you guessed it, cancer biology. Their result after 8 years of effort? A success rate of 46%.
More recently, in February 2026 in France, Raphaël Lévy (Sorbonne Paris Nord University) and his team published the first results of their NanoBubbles Replication Initiative. The project aims to replicate results of influential nanobiology studies (which propogate as “information bubbles”). Spoiler: their first attempt was abandoned midway, as the replication results differed significantly from the originals. But the team is continuing with other papers!
On a larger scale, the University of Bern conducted an experiment called “ERЯOR,” which paid scientists to find errors in publications. The initiative appears to have been less successful than paying reviewers, and was recently discontinued in favour of a scientific journal for articles reevaluating or replicating previous work. By offering publication opportunities instead of cash rewards, this new phase of ERЯOR hopes to better recognize this labour-intensive work.
One last example: a hackathon at the International Conference on Machine Learning in August 2026 challenged people to reproduce the results of as many papers as possible. A cybersecurity agent (not a scientist) won, reproducing the results of 363 scientific papers in 19 days with the help of his AI agents. Impressive, but such success was only possible because of the papers’ well-documented and accessible data and methods. The achievement, however, also raises an interesting question: rather than optimizing AI for peer review itself (a growing practice which remains controversial), should we be optimizing AI tools to test articles’ computational reproducibility as part of peer review evaluation?
Approach #2: Adopt the best practices to make analyses reproducible from ‘go’
Reproducibility at hackathon scale remains, for now, the preserve of purely computational fields. Yet many other scientific disciplines rely heavily on computational analysis, and all deserve the same level of transparency and reproducibility. A recent article offers sound advice for all of us. It acknowledges that making an analysis pipeline reproducible is time-consuming and that, for ethical reasons, it is not always possible to share data openly. However, the authors defend the long-term value of transparent-as-possible pipelines: this increases the likelihood that your work will be adopted and reused elsewhere, thereby boosting your visibility and citations. Even small efforts can make a major difference.
The article includes a “checklist”; here are its most important and/or original suggestions:
- Write organized and clear README files, including details about each environment (starting with the local computer) and the exact version of each tool or package used.
- Organize and name your code, files, and paths in a logical and easy-to-follow manner.
- Write separate code files to reproduce each analysis and figure, matching their numbering in your manuscript.
- Set and specify random seeds any time there is a calculation involving a random element.
- If sharing real data is not possible, provide synthetic data to enable testing and validation of your code (and explain why the actual data cannot be made openly available).
Of course, following all of this advice involves a considerable amount of extra work, particularly the creation (and evaluation before sharing) of synthetic data. However, as someone working in the field of human genetics which often includes senstive data, I think using synthetic data to facilitate code evaluation could be a very elegant solution.
In the short term, the best strategy will surely be to change the way we work, structuring our research so that it is as reproducible as possible from the very beginning. This will make it easier to submit papers that adhere to these practices, and we’ll also save time, since it will be much easier to reproduce, adapt, and reuse our own analyses. Good luck!
Caitlin Martin, postdoctoral researcher at the Institut Pasteur
References:
- Haven TL, Ioannidis JPA. Reproducibility Failure in Biomedical Research: Problems and Solutions. Annu Rev Med. 2026;77:537‑49. doi:10.1146/annurev-med-050124-050859
- Errington TM, Mathur M, Soderberg CK, Denis A, Perfito N, Iorns E, et al. Investigating the replicability of preclinical cancer biology. Pasqualini R, Franco E, éditeurs. eLife. 7 déc 2021;10:e71601. doi:10.7554/eLife.71601
- Said M, Gharib M, Lévy R. Towards More Robust Nanobioscience: Launching the NanoBubbles Replication Initiative. Angew Chem Int Ed. 2026;65(12):e11631. doi:10.1002/anie.202511631
- Hornung R, Németh L, Zadorozhnyi O, Ullmann T, Kammer M, Killick R, et al. Overcoming barriers to computational reproducibility. R Soc Open Sci. 1 juin 2026;13(6):252489. doi:10.1098/rsos.252489
