Services · Protein models
Reveal hidden features of your protein language model.
We break a protein language model’s internals into individual features and match each one to the shape of backbone it responds to. You get a map of what the model represents, feature by feature, before you build on it.
Read the paper · NeurIPS 2026 ↗Unrelated proteins, one shared turn of backbone, matched to 0.3 Å. A feature keyed to the shape reads both; a database sees two different proteins.
Validate
Before trusting a binding-site prediction, check the model keys on the pocket’s shape, not a shortcut in the sequence.
Discover
Spot a pocket shape the model links across unrelated protein families: a new drug target, or an off-target risk, before the wet lab.
Screen
A model that designs proteins can design toxins. See whether its features respond to toxin folds before the weights go out.
Building or serving a biological foundation model?
Tell us which model, and what you need to know about it.
Edinburgh, Scotland