cradle
cradle.bio/blog/
unfold-wrap-up
Case study
Four things protein engineers taught us at our first user meeting
A recap of Unfold 2026

Christine

On November 4, several of my colleagues and I spent an afternoon at View Boston with scientists who use Cradle, and scientists who are thinking about it. We built the format around Q&As, roundtables, and breakouts because we wanted scientists talking to each other, not to a slide deck.
“Built for scientists” means letting scientists help shape what comes next. Here are four things we took away from those conversations.
1-Scientists want to understand what informs model suggestions
Johan, a Cradle Scientific Solutions Consultant based in Amherst, MA, who shows scientists how to use Cradle to build better proteins, faster, walked a small group through the UI, with Aviv (same role, Boston-based) fielding questions. Within a few minutes the questions had stopped being about the interface and started being about the model’s knowledge of biology.
Does it only handle certain assay types? Do I need a structure or epitope data? If I have prior knowledge of certain residues, can I block mutations there? Do the models understand what a mutation does near a structural motif, given that nobody hands it a structure? With only a handful of data points, can they pick up epistasis?

Those are good questions and they share a premise: A scientist is being asked to trust suggestions from a system whose reasoning they can’t inspect the way they’d inspect a docking result. People want to know what the model is generalizing from before they spend a plate on it.
We’ve published on how to read trade-offs in multi-property optimization and on how we benchmark, and a plain-English explainer on protein language models is coming shortly. But we have work to do here inside the platform, too.
2-Teams build custom automations (and reduce clicking)
One attendee presented how his group rebuilt its interaction with Cradle around the API. They wrapped the workflow in a command-line interface driven by a single versioned config file per project, and set it up so each step kicks off the next one automatically instead of waiting for someone to notice the previous one had finished. Work that used to take between two and thirteen hours per cycle now takes a few minutes, and none of what was removed was science. It was downloading sequences, converting protein to DNA, building constructs, and normalizing assay data by hand.
They kept a human at the build step, because that is where you commit thousands of dollars of DNA synthesis. The clicking disappeared, and scientist decision-making stayed.
Taking Cradle in this direction meant adapting a few things on our side. Our login flow was built around a person at a keyboard, which doesn’t suit a pipeline that runs overnight. We’ve since added machine-to-machine token support, so automated runs no longer get interrupted.
3-Service is great, doing things yourself is better
Our Scientific Advisors got a lot of praise at this event. Sometimes called “forward-deployed scientists,” these are the Cradlers who help customers with their programs; it’s a team that is very hard to build. Several people also made it clear they’d rather not need them.
It comes down to setting up a round. The first time anyone sits down to configure a project, the sensible defaults often still require guidance through an expert. With molecules and scientific challenges being very specific and different per project, it's not always obvious where to start.
We're excited that scientists using Cradle want to work independently, and we're constantly iterating on our platform to make this easier. We've recently launched an all-new Cradle Academy for our clients with step-by-step guides, but the work is never finished.
4-Every customer’s data looks different
Every customer’s data looks different. Different LIMS, different assay naming, different conventions for what a plate identifier means. We’ve added a lot of functionality to make formatting and solving data issues easier, but it’s still not always smooth sailing to get data into a state that works with Cradle.
One customer had spent significant time turning Genedata exports into something our upload would accept: distributing merged header cells, concatenating names to keep them unique, stripping characters, converting symbols into words. On upload, issues weren’t immediately clear.
We have documentation for data prep available to customers, and a new Cradle Academy data-screener agent that automates most of that cleanup, because no one should be spending a week doing this. We’re matching our label-length limits to what Genedata produces, making upload more permissive about merged headers and special characters, allowing amendments instead of full re-uploads, and adding provenance to downloads so you can see which template a result came from. (On that last point, our post on sample tracking covers why lineage matters more than it first appears.)

What I’m taking back
When you build a platform, you learn your own product too well. You know where to click, what a term means, why a decision was made. The scientist using it shouldn’t need that background to get value from it.
“Built for scientists” means taking those experiences to heart and letting them shape what we build next. Our first user group meeting brought that principle to life: scientists sharing how they work, learning from one another, and helping us understand what would make the biggest difference in their day.
Unfold was also the beginning of a community we’re excited to keep bringing together. Keep showing us what you’re working on, where you get stuck, and what you’d like to try next. We’re taking that feedback back to our teams as we build software that lets scientists put the latest advances in AI to work. We look forward to sharing the progress and hearing what comes next when we meet again. See you at Unfold 2027!
Cradle
cradle.bio/blog/
unfold-wrap-up