There's a huge disconnect in many life sciences labs. There are computational teams building complex models and wet lab scientists doing research at the bench. These two worlds don't always mix. This means the scientists who run experiments often get stuck with tedious, manual digital work, tasks that feel like they should have been automated a decade ago. They rely on machines for data but often lack the programming background to build the custom tools they desperately need.
A concrete example from mRNA therapeutics is selecting the right gene variant for drug screening. Before a new drug can be synthesized, a researcher must choose the correct variant of a target gene (isoforms). TP53 for example has 12. Picking the wrong one can mean targeting the wrong tissue or expressing a non-functional protein, leading to false negatives that can kill a promising therapeutic before it even gets a shot.
The manual workflow followed to get relevant data for any target follows a process I imagine many recognize, even if the science is different:
1. Navigate through several database websites that look like they were designed in 1999.
2. Click through menus to locate the correct string of letters buried in a sea of plain text.
3. Manually copy-paste a long sequence of text into a spreadsheet.
4. Ask a colleague to double check this for however many targets you have. They will then, reproduce (or pretend to reproduce) the same mind-numbing steps as you.
The whole process is slow, incredibly error-prone, and a small oversight can derail weeks of research. The scientific choice also carries real consequences, especially when a researcher is selecting between multiple isoforms of a single gene. Yet the process often relies on vague, unstated heuristics; for example, “just pick the longest one” when a dozen transcript variants are returned. Hundreds, if not thousands, of hours are wasted in similar manual ways every year across research teams in the U.S.
However, vibe coding is proving to be one of the greatest gifts to experimental scientists who are curious enough to try. Emerging systems make it possible for anyone to write quick, bespoke, good-enough scripts that save hours of repetitive work. For example, I used this approach to build a simple pipeline that automates this entire tedious process.
The repo with the case study is here: Repo
Substack article is here: Vibe Coding, mRNA, and a Very Conflicted AI Claude
When I read discussions from software engineers who are skeptical of vibe coding, I can’t help but think they’re blinded by their own experience. They either can’t fathom , or have long forgotten, the tremendous acceleration and excitement that occurs when you leap from a manual standstill to modest automation. My suspicion is that domains with a natural tolerance for “shooting fast and from the hip” will gain the most from the lightweight automation these tools now make possible.
github repo here: https://github.com/gvmfhy/constitutional-seq
substack post here: https://austinpatrick.substack.com/p/rapid-mrna-drug-design-...
