Rowan Docs

Protein Binder Design

Python API · stjames models · API example

How it works

Rowan uses BoltzGen to propose protein, peptide, or nanobody binders for a specified target. The pipeline generates candidate structures and sequences, evaluates how the sequences fold with the target, then filters and selects designs using predicted quality and structural diversity. Results include designed sequences, predicted bound complexes, and quality scores.

Settings

  • Target and binder: supply target protein structures or ligand SMILES and specify the binder sequence pattern. Each chain needs a unique ID. At least one sequence must contain a designable region: 80..140 requests 80–140 new residues; 9A9C designs nine residues, fixes an alanine, designs nine more, then fixes a cysteine. Letters alone specify a fixed sequence. Peptides can also be cyclic.
  • Structure regions: "Include regions" and "Exclude regions" select which chains and residues form the target context; all are included when no selection is specified. "Binding regions" specifies where the binder should or should not contact the target. These settings guide design and should be checked in the resulting complexes.
  • Protocol: "Protocol" selects protein, peptide, or nanobody design, or design of proteins that bind small molecules. The web form defaults to protein design (protein-anything).
  • Candidate count: "Num designs" sets how many candidates to generate; the web form defaults to 10. Larger counts increase the search effort and runtime. Start with a small run to inspect the design specification before scaling up.
  • Final selection: "Budget" requests the number of designs in the final selection after quality filtering and diversity selection; the web form defaults to 2. Use a positive value no larger than "Num designs."

Notes

Compare predicted interfaces and sequences alongside confidence, backbone agreement, interface contacts, and sequence liability scores. Available metrics vary by protocol. A high confidence score or favorable interface does not establish experimental binding, specificity, or expression; selected designs require experimental validation.

The Python API also supports proximity selections, secondary-structure preferences, redesign regions, insertions, and covalent-bond constraints. These advanced controls are described by the linked stjames models.

Further reading