Practical bioinformatics tutorials and notes, from bench to insights.https://benchtoinsights.com/enhttps://benchtoinsights.com/blog/reproducibility-in-bioinformatics/https://benchtoinsights.com/blog/reproducibility-in-bioinformatics/A few years ago I tried to reproduce a figure from one of my own papers, and it would not run. Here is what that taught me about reproducibility.Tue, 07 Jul 2026 00:00:00 GMT<p>A few years ago I tried to reproduce a figure from one of my own papers. The script was there. The data was there. It still would not run. R had moved on, two packages had changed their defaults, and the reference genome I had used was a build I could no longer find on my drive. It took me most of a day to get back to a plot I had first made in an afternoon. That is the moment reproducibility stops being a buzzword and becomes your problem.</p> <p>Bioinformatics is software. We forget this. A wet lab protocol lists the antibody clone, the lot number, the incubation time. Our version of that is the aligner and its version, the genome build, the random seed, and the parameters we typed at 2am and never wrote down. Leave any of them out and the analysis turns into a story instead of a method.</p> <p>The literature has said this for years. Sandve and colleagues published their ten simple rules for reproducible computational research in 2013, and the list still holds. Peng argued in Science in 2011 that reproducibility sits between full replication and nothing at all, and that computational work has no excuse for landing on nothing. When Nature surveyed researchers in 2016, most admitted they had failed to reproduce another group’s results, and many had failed to reproduce their own. None of this is new. We just keep relearning it.</p> <p>In practice it comes down to habits, not good intentions.</p> <ul> <li>Pin the environment. conda, renv, or a container. Record versions, not just names.</li> <li>Kill hardcoded paths. A config file and a sane project layout beat a home directory full of absolute paths.</li> <li>Set the seed and save your sessionInfo. Future you will want both.</li> <li>Let a workflow manager drive the pipeline. Nextflow and Snakemake exist so you do not have to remember the order of forty steps.</li> </ul> <p>I reach for Nextflow and nf-core for anything I expect to run more than once. Di Tommaso and colleagues built Nextflow around this exact problem, and Köster and Rahmann had done the same with Snakemake years earlier. Containers handle the unglamorous part and freeze the operating system underneath. Put those pieces together and a stranger can clone the repo, run one command, and land on your numbers. That is the bar.</p> <p>None of this makes the science correct. A reproducible pipeline can be reproducibly wrong. But you cannot debug, review, or trust an analysis you cannot run twice. Reproducibility is what turns a throwaway analysis into something you can stand behind. It is the cheapest insurance in the field, and we still underbuy it.</p> <div><h2 id="further-reading">Further reading</h2></div> <ul> <li>Sandve, Nekrutenko, Taylor, Hovig. Ten Simple Rules for Reproducible Computational Research. PLoS Computational Biology, 2013.</li> <li>Peng. Reproducible Research in Computational Science. Science, 2011.</li> <li>Baker. 1,500 scientists lift the lid on reproducibility. Nature, 2016.</li> <li>Di Tommaso et al. Nextflow enables reproducible computational workflows. Nature Biotechnology, 2017.</li> <li>Köster, Rahmann. Snakemake, a scalable bioinformatics workflow engine. Bioinformatics, 2012.</li> </ul> reproducibilityworkflows