name: esm2-guided-mutation description: Generate one-step CDR mutation proposals from ESM-2 deep-mutational-scan log-likelihood ratios, then pass those proposals to the normal fold and gate pipeline. metadata: version: "1.0.0"

ESM-2 Guided Mutation

Use this as the primary proposal skill only when the parent is a complete, materialized sequence and the Router exposes it.

The runtime performs the ESM-2 DMS and materializes the highest positive LLR substitutions at mutable CDR positions. The Design Agent must still select this skill explicitly and provide a valid top-level portfolio. ESM-2 is a sequence plausibility model, not an affinity predictor.

Output contract

The Agent only selects this Python-executed skill. Return exactly one JSON object with no prose or code fence before or after it:

{
  "skill_id": "esm2-guided-mutation",
  "selection_reason": "ESM-2 sequence-plausibility rescue best matches the current bottleneck.",
  "candidates": []
}

Do not invent mutations or placeholder candidates; Python performs the DMS and materializes the proposals.

Contract

• Python proposes substitutions only at the listed mutable positions using the

configured semantic chain IDs. Agent-facing sequence positions are one-based; the runtime uses zero-based indices internally.

• The runtime records esm2_llr, the parent residue, and the selected rank in

candidate metadata before folding.

Guardrails

• Never use this skill for a parent containing X or an unmaterialized CDR;

full redesign must materialize the sequence first.

• A positive LLR does not imply better antigen binding. Keep OpenDDE metrics,

contact-fraction gates, and developability checks authoritative.

• Treat the generated mutations as proposals; all candidates still undergo

folding, admission, and population selection.