aurekaresearch/opendde-harness/minibinder-inverse-folding

Mini binder inverse folding

Requires the actual parent structure path supplied in context. Select this skill when backbone-conditioned sequence sampling is appropriate. Never invent a structure path or reconstruct it from a filename.

Return one anchor candidate with mutations: [] and optional metadata.soluble_mpnn_parameters. Supported controls are temperature (0.01–1), relax_radius (0–8), wt_bias (0–20), omit_aas, bias_aas, and design_positions (chain to one-based position list). Omit controls when no evidence supports changing their defaults. Do not set num_sequences: Python uses the configured cycle candidate count, retries missing unique samples for at most three sampling rounds, and reports an error if the batch remains incomplete.

Design positions must be a subset of the supplied mutable mask. The executor owns chain mapping, fixed residues, chain length, deduplication and total candidate count. It generates sequences; do not claim that generation or validation already happened. Re-fold and evaluate candidates before interpreting them as improvements. This changes sequence on an existing backbone, not the backbone topology.