Biomaterials Testing Barely Changed in 30 Years

Biomaterials testing is still mostly done by hand.

Walk into a drug discovery lab and the screening runs itself. Robots move liquids across 384 and 1,536-well plates, around the clock, throwing off millions of clean data points.

Walk into a hydrogel lab and you’ll find a scientist with tweezers and a balance.

A 2024 University of Rostock swelling study captures it perfectly. To measure how the gels took up water, researchers pulled each sample out by hand, blotted it, weighed it, put it back, and repeated that at set intervals for the length of the study.

That’s roughly how it was done 30 years ago.

And this is the part people miss about AI. It didn’t transform drug discovery because someone wrote a clever model. It transformed it because automation had already produced mountains of structured data for a model to learn from.

Biomaterials never got that automation.

So there’s no data.

So there’s nothing for AI to learn.

Any model we build is only as good as the measurements underneath it. Right now, those measurements still start with a person in the lab.

A lunch in Boston

When I was a student, I went to a pharma conference in Boston.

A senior scientist took me out to lunch. He’d spent 20 years in the field, had a PhD, and was working on an injectable product at one of the biggest pharma companies in the world.

I told him I’d started building robots because I was tired of feeling like one. My days were some version of weigh, swell, blot, reweigh, record, repeat.

He laughed. He was doing the exact same thing.

I’d assumed a company that size had quietly solved this years ago. They hadn’t.

That lunch changed how I saw the problem. It wasn’t my lab. The tools were missing across the whole industry.

Why gels break normal automation

Almost all lab automation was built to move liquids.

Drugs, DNA, and small molecules move fine with a pipette. They behave predictably.

Hydrogels don’t. They’re soft solids made mostly of water. A sample the size of a grain of rice can soak up hundreds of times its weight and change in size and stiffness while you’re still testing it.

Try to grab one with a spatula and it can tear in half. Try to pipette it and the viscosity wrecks your accuracy. A liquid handler can’t treat a gel like another well on a plate.

So the test stays manual:

Dry the sample

Drop it in buffer

Pull it out on a schedule

Blot, weigh, repeat

Run that across dozens of formulations for days, and two things happen. The work buries whoever’s doing it. And a material that changes continuously gets flattened into a handful of time points.

We’re trying to understand dynamic materials with static tools.

AI needs the measurements first

NIST made this point in a 2024 paper on polymer characterization.

Useful AI depends on three things stacked together:

high-quality data,

automated experiments,

and shared standards that connect the hardware, the software, and the data.

Hydrogels fail all three.

Results sit in notebooks and spreadsheets. Every lab runs slightly different conditions. Two people following the same protocol can hand you two different swelling curves.

AI can compare formulations and tell you what to run next. It can’t recover a measurement nobody took, and it can’t turn messy testing into a reliable prediction.

The physical work has to come first. Handle the gels without wrecking them. Automate the swelling and degradation studies. Capture continuous data under controlled conditions.

Get that right, and the models finally have something to learn from.

The layer nobody wanted to build

Most people in biomaterials want to invent a new gel, or put a new drug inside one.

I get it. The new material is the exciting part. It’s the thing with your name on the patent.

The testing underneath it is nobody’s dream project.

So that’s where we started.

At GelTech Labs, our hardest problem was learning to handle the gels themselves without destroying them.

Once we cracked that, Carousel could run six formulations at once, around the clock, across a range of pH and temperature, logging a continuous curve instead of a few scattered points.

The scientist stops shuttling every sample between a buffer and a balance. The data comes out consistent enough to actually train on.

Every new material in this field eventually hits the same testing wall. If we don’t fix that shared layer, everything built on top of it moves slower than it should.

Hydrogels are 21st-century materials. We still test them like it’s 1995.

Fix that, and the models finally get data worth learning from.

Then, the materials take care of themselves.

First published on LinkedIn.