Structure
Model
Adds the side-chain atoms of nearby residues you are not designing to the atom context. Because it reads the selection, changing which residues are designed re-runs the encoder.
Label the residues selected in the viewer. Untouched residues stay soluble.
Design
Low is conservative and repeatable; high explores.
Bias & omissions
Bias is added to the logits before sampling, so +2 makes a residue roughly e² times more likely — it only wins where the model's own preference is weaker than that. Omit removes a letter outright.
Tied positions
Tied positions are decoded together and share one identity. Their logits
are averaged, so a group behaves like a single position; the
sum mode adds them instead, which sharpens as the group grows.
Credits and licences
Weights and model architecture from dauparas/LigandMPNN (MIT). Speed and API ideas from ColabDesign; page structure after CIRPIN-web. Structures fetched from RCSB and the AlphaFold DB.
Sequence
Saturated is selected · drag to sweep · click a chain badge for the whole chain
Results
Designs will appear here. Click one to paint it onto the structure.