aurekaresearch/opendde-harness/antibody-inverse-folding

Antibody Inverse Folding

Use this as the only primary skill when the Router selects the atomic antibody-inverse-folding skill. Router-approved context providers may also be loaded, but no other primary design skill may be composed with it.

1. Require a valid parent/backbone structure, binder chain mapping, parent

sequence, and mutable/fixed positions. Never invent missing chain mappings.

1. Propose explicit, diverse CDR anchor substitutions. For every anchor, choose

SolubleMPNN sampling controls in metadata.soluble_mpnn_parameters based on the current bottleneck instead of assuming one fixed setting: temperature, relax_radius, wt_bias, omit_aas, and optional global bias_aas. To choose an exact CDR subset, provide design_positions as a chain-to-one-based-position object; it is always intersected with the authoritative mutable CDR list. Python validates these values, allocates the configured candidate count across the anchor plans, hard-fixes each forced anchor, and never permits framework positions to become designable.

1. Treat SolubleMPNN as the sequence generator, not as evidence that a candidate

binds. Preserve the strategy skill's hypothesis and let downstream folding, CDR-RMSD, interface, and developability gates evaluate the sample.

1. Keep chain identities and one-based sequence-position semantics consistent

when translating to structure residue numbering.

Output contract

Return exactly one JSON object with no prose or code fence before or after it. Use the same candidate JSON shape as every other primary design skill:

{
  "skill_id": "antibody-inverse-folding",
  "selection_reason": "Structure-conditioned local resampling best matches the current bottleneck.",
  "candidates": [
    {
      "id": "cycle_N_inverse_001",
      "mutations": [["D", 104, "F"]],
      "strategy": "[H3 / inverse-folding] structure-conditioned local redesign",
      "risk_level": "medium",
      "metadata": {
        "anchor_rationale": "stabilize the H3 tip",
        "soluble_mpnn_parameters": {
          "temperature": 0.25,
          "relax_radius": 3,
          "wt_bias": 2.0,
          "omit_aas": "",
          "bias_aas": {"Y": 0.5},
          "design_positions": {"D": [102, 103, 104]}
        }
      }
    }
  ]
}

Pipeline rules

• Do not execute shell commands or fabricate SolubleMPNN scores. The Python runner

owns model execution through generate_soluble_mpnn and the opendde_harness.plugin.protein_design.servers.backends.soluble_mpnn backend. This uses LigandMPNN's soluble_mpnn model, not its ligand-conditioned model. SolubleMPNN scores are mean negative log probabilities over designed residues (lower is better); confidence is exp(-score) (higher is better).

• Return mutation-style JSON anchors under top-level

"skill_id": "antibody-inverse-folding". The runner uses them as independent structure-conditioned generation hypotheses.

• Return at most the requested number of anchor plans. Python distributes the

cycle's configured candidate count across all plans; do not set num_sequences. It keeps valid unique sequences and samples missing candidates for at most three rounds. A remaining shortfall is reported as an error, not a completed batch.

• Encode every anchor in the canonical one-based three-item array form

["D", 104, "F"] inside "mutations". Do not emit mutation objects with chain, position, from, or to keys, and do not emit compact strings such as D104F. The runner accepts those forms only as compatibility fallbacks.

• Context providers are automatic evidence sources; do not declare them in candidates.
• Preserve fixed_residues, design_method, and chain mapping in generated

candidate metadata.

• Put skill-specific proposal evidence only in the metadata JSON object.
• If the required structure or SolubleMPNN client is unavailable, report that the

inverse-folding route cannot execute; do not silently relabel an LLM sequence as inverse-folded.