Making a protein that sticks to a chosen target is hard because the space of possible amino acid sequences is enormous and testing each one is expensive.
This work treats design as a search over sequences: a structure-predicting neural network suggests which edits to try, and a full run of that network scores each candidate. Early rounds edit several positions at once to explore; later rounds adjust one position at a time.
In tests against several targets, designs scored better and more often passed strict confidence checks than baselines. In the lab, 82.9% of the small proteins designed against one target bound it, the strongest with a binding affinity of 0.879 nanomolar, where lower is tighter.