Abstract: Everyone is excited about AI for drug discovery. Nobody is talking about the fact that the data needed to make it work is locked in desktop app files on a shared drive.
There’s a pattern in every serious AI for biology conversation: the model is smart, the benchmark numbers are impressive, and then someone asks “how do you get the data in?” and the conversation gets quiet.
The dirty secret of AI for biotech is that most lab data isn’t queryable. It’s in FlowJo .wsp files, Prism .pzfx files, and Excel spreadsheets on a network drive organized by whoever joined the lab three years ago and quit in 2023.
What “AI-ready data” actually means
AI-ready doesn’t mean “digital.” Most lab data is digital. It means:
- Structured: a query can retrieve it without knowing which file to open
- Labeled: the model can distinguish between a negative control and a treatment replicate
- Lineage-tracked: you know what version of the protocol produced which result
- Accessible: the AI system has a governed path to the data, not standing credentials to a shared drive
Most lab data fails all four of these. It’s digital, but it’s locked in application-specific binary formats that require the original software to open. It’s labeled only in the scientist’s head and a notebook entry three columns over. The lineage is “ask the postdoc.” The access pattern is “here’s the password to the VPN.”
Why top-down harmonization fails
The standard enterprise answer to this problem is a harmonization project: hire a data engineering team, build a pipeline from every instrument into a central warehouse, and normalize it after the fact.
This fails for two reasons.
First, the normalization is lossy. When you extract “the Ct values” from a qPCR run and put them in a table, you’ve discarded the amplification curves, the raw fluorescence, and the analysis parameters that produced those Ct values. If the analysis was wrong, you can’t go back.
Second, scientists don’t change their workflow for the data team. They keep using FlowJo and Prism and pasting summaries into Benchling, and the pipeline is always six months behind the science.
The right architecture: capture at the source
The only way to get clean, structured, lineage-tracked lab data is to be the tool the scientist actually uses to do the analysis. Not a wrapper around it — the thing itself.
If the analysis happens in a system that knows the raw file, the experiment, and the analysis parameters, then the structured data and the lineage are both a side effect of the scientist’s normal daily work. Nothing extra to do. No harmonization required.
This is the core idea behind EpistaBase: build the tool that’s better enough for scientists to choose it for daily work, and the governed, queryable data layer fills itself.
The AI copilot that runs against real lab data — instrument files, experiments, statistical results — rather than a harmonized summary is qualitatively different from one that runs against extracted CSVs. It can answer questions like “what’s the trend in binding affinity for our mAb variants above pH 7?” against the actual underlying data, not an incomplete extraction of it.
That’s the data layer biotech has been missing. It won’t come from a data engineering project. It’ll come from scientists choosing to use better software.