The model learns what a plausible protein looks like. Still, this data covers less than 0.01% of all natural sequences.

Our wet lab in Amsterdam measures how sequence changes affect expression, stability and solubility. Public databases do not have this data.

The model trains again on sequences from your protein family. It learns which positions you can change, and which positions you must keep.

Assay data from your company and this project. The model learns all your properties and their trade-offs.

The model designs the next variants. The lab tests them, and the model is retrained on all results so far. A few rounds are usually enough.