Hey everyone, it’s Tommy here. If you’ve been following my blog or my Twitter/X (@tangming2005), you know I love diving into the practical side of bioinformatics and genomics.
Recently, I gave a talk titled “Good Enough Practices for Reproducible Computing” at Moderna, where I spent a good chunk of time chatting about reproducible computing.
Why? Because in our field, where data is exploding and analyses get complex, making sure your work can be repeated—by you or anyone else—is a game-changer. I thought it’d be fun to turn those slides into a blog post here.
In this post, I’ll focus on the reproducible computing part, sharing why it matters, why it’s tricky, and some simple tips to make it happen. I’ll keep it straightforward, like I do in my other posts—no fancy jargon without explanation. Let’s jump in!
Why Bother with Reproducibility?
First off, remember that story from Keith Baggerly at MD Anderson? He uncovered issues in high-throughput biology papers because methods weren’t clear or reproducible. Check out his hilarious YouTube talk on it: The Importance of Reproducible Research in High-Throughput Biology.
Stuff like that shows how non-reproducible work can lead to wrong conclusions, wasted time, and even bad decisions in drug development—which can be super expensive.
But here’s a personal angle: Your closest collaborator is yourself six months from now. I’ve been there—digging up an old analysis and thinking, “What the heck did I do here?”
If it’s not reproducible, you’re starting from scratch. And in biotech/pharama, where I’m now Director of Bioinformatics at AstraZeneca, we can’t afford that. Reproducibility saves time, builds trust, lets others build on your work and most importantly, it saves lives.
Titus Brown summed it up nicely in one of his talks: Reproducibility isn’t just nice; it’s essential in computational biology.
The Challenges: Why Is It So Hard?
From my experience (and the gaps in my own career journey from wet lab to comp bio), here are the big hurdles:
- Data and Scripts Go Missing: Raw data isn’t versioned, or scripts are “available upon request” (yeah, right).
- Vague Methods: Papers skip details on tools, versions, or even operating systems.
- Version Mismatches: Different R/Python packages, bioinformatics tools, or OS (Mac vs. Linux vs. Windows) can change results.
- No Standards: Everyone organizes projects differently, making collaboration a nightmare.
I’ve seen this in my over a decade long career in bioiinformatics. One mismatch, and poof—results don’t match.
How to Make Your Work Reproducible: Practical Tips
The good news? You don’t need rocket science. Here’s what I covered in the talk—simple steps:
Version Your Data and Files Smartly
Version control large files with git lfs.
Name files like a pro: Use ISO 8601 dates (YYYY-MM-DD), no spaces or special chars, and add slugs for clarity. For example, “2025-08-17_reproducible-analysis_results.csv” beats “final_v2.csv”.
Shoutout to Jenny Bryan for her awesome slides on this: Naming Things.
Organize projects consistently: Folders like /data (read-only), /scripts, /results. In R, use
here::here()for paths—avoids headaches. Python folks, check out pyhere on GitHub.Git for Code Versioning
Git is your best friend for tracking changes. I commit often and push daily. Basic commands:git init,git add,git commit,git push.It saved me when inheriting projects at Immunitas Therapeutics (after people left). If you’re new, try Happy Git with R or Learn Git Branching.
Pro tip: Use branches for experiments you might toss.
Manage Environments with Tools Like Conda, uv, or renv
Pin package versions! For Python, I love mamba (faster Conda) or uv (super quick, check it out on GitHub: https://github.com/astral-sh/uv).In R, renv snapshots your library:
renv::init(),renv::snapshot(). No more “It works on my machine” excuses.Containers for the Win: Docker or Singularity
Think of Docker as a virtual machine that bundles everything—OS, packages, code. I’ve used it for ChIP-seq pipelines. Resources:Rocker Project for R in Docker,
BioContainers for bio tools.
Literate Programming: Mix Code and Words
Use Jupyter Notebooks for Python or R Markdown/Quarto for R. I will need to switch to Quarto—it’s next-gen R Markdown (https://quarto.org). Embed code chunks, explanations, and outputs.Bonus: Quarto is Git-friendly. Document outside too, like in my enhancer-promoter repo: https://gitlab.com/tangming2005/Enhancer_promoter_interaction_data.
Python users want to take a look at marimo as a Jupyter alternative
Automate Everything
Scripts over manual tweaks! For repetitive tasks, bash scripts or workflows like Snakemake/Nextflow. In R, {targets} is gold (book: https://books.ropensci.org/targets/).
Automation = best documentation. Computers love boring work; let them handle it.Clean Code with Functions
Don’t repeat yourself—use functions. In R, purrr::map() for loops. I even ask ChatGPT to refactor messy code. For bigger stuff, build packages with roxygen2 (book: https://r-pkgs.org).Functional programming keeps things tidy and reusable.
Good Enough Practices to Get Started
You don’t need perfection. Start with:
- Consistent folders.
- Notebooks for analyses.
- Extensive documentation.
- Git for versioning.
- A quick HTML report with knitr.
- End every project with a slide deck linking to your GitHub.
Over time, aim for the full spectrum: Reproduce your own work anytime, anywhere, and let others do the same.
Tools that are useful
SciDataFlow — Facilitating the Flow of Data in Science. By Vince Buffalo, author of Bioinformatics data skills (one my favorite books).
pracpac: Practical R Packaging with Docker. By Stephen Turner.
Shournal A (file-) journal for your shell.
I do not want to write down every linux commands in the terminal anymore. Check out Shell Sync
Script of Scripts (SoS) is a computational environment for the development and execution of scripts in multiple languages for daily computational research. It can be used to develop scripts to analyze data interactively in a Jupyter environment, and, with minimal effort, convert the scripts to a workflow that analyzes a large amount of data in batch mode.
DSO: is a command line helper for building reproducible data anlaysis projects with ease from Boehringer-Ingelheim.
Imagine being able to click on a plot and seeing the complete Jupyter/Rmd notebook, data, and parameters. That’s GoFigr!
Two books I have read
Building reproducible analytical pipelines with R. Highly recommend if you use R.
Software Engineering for Data Scientists: From Notebooks to Scalable Systems if you use python. But the concepts are applicable no matter what language you use.
Wrapping Up
Reproducible computing transformed my career—from PhD at UF to leading teams now. It avoids errors, speeds things up, and makes you a better scientist.
If you’re in bioinformatics, embrace it! Got questions? Hit me up on Twitter or in the comments below. And if you like this, subscribe to my newsletter for more tips: https://divingintogeneticsandgenomics.ck.page/profile.
What do you think— what’s your biggest reproducibility headache? Let’s chat!
(Oh, and if you’re into videos, check my YouTube channel Chatomics for related tutorials.)