molclaw-boltz2-affinity

Files (1)

• SKILL.md

SKILL.md (first lines)

# Boltz-2 Protein-Ligand Binding

## Invocation

Use the exact tool name and parameter schema exposed by the current harness. Tool names in SDK examples below are raw MCP names, not model-facing aliases. Local files must be uploaded with `molclaw-file-transfer` before a server tool can use them. Repair PDB inputs first when the workflow requires it.

step 1. Use skill **molclaw-protein-sequence-retrieve** to get the target protein sequence information. If the target protein sequence has been provided, skip this step.

step 2. Finally use tool *pred_binding_affinity_boltz2* to predict the binding affinity.  

Tool description:

```tex
Use Boltz to predict binding affinity between protein (receptor) and small molecule (ligand).
The server selects the output directory. This tool is for small-molecule ligands, not peptide/protein partners; ligands exceeding the Boltz affinity atom limit are returned as a structured model-capability error rather than a timeout.
Args:
    protein (List[dict]): Protein chains, each element contains 'chain' and 'sequence' (e.g., [{{'chain': 'A', 'sequence': 'MGNAAAAKKGSEQASQRRSSLEQP*'}}])
    smiles (str): Input SMILES string (e.g., "N[C@@H](Cc1ccc(O)cc1)C(=O)O")
Return:
    status (str): success/error
    msg (str): message
    affinity_probability_binary (float): Represents the predicted probability (ranging from 0 to 1) that a ligand is a binder, making it ideal for distinguishing active compounds from decoys during the hit-discovery stage. A value below 0.5 indicates uncertain or weak binding.
    affinity_pred_value (float): Estimates the specific binding affinity as log10(IC50) in μM to quantify how small molecular modifications affect potency, serving as a key metric for ligand optimization phases like hit-to-lead and lead-optimization.
    complex_cif_file (str): Structure file of the protein–molecule complex
```

Tool usage:

```python
response = await client.session.call_tool(
    "pred_binding_affinity_boltz2",
    arguments={
        "protein": protein_chains,
        "smiles": smiles
    }
)
result = client.parse_result(response)
affinity_probability_binary = result["affinity_probability_binary"]
affinity_pred_value = result["affinity_pred_value"]
```

Current capability boundary: Boltz affinity rejects ligands with more than 128 atoms. For peptide ligands such as PTHrP/TIP39 fragments, this is expected behavior; use protein-peptide structure/docking workflows such as Chai-1/HDOCK plus `interaction_visualizer(mode="peptide")` instead of interpreting the Boltz rejection as a server failure.

From InternScience/MolClaw, MIT.

More from InternScience/MolClaw