dgx-spark-playbooks/nvidia/station-healthcare-agent/assets/skills/molecular-viz/SKILL.md

name	description	metadata
molecular-viz	Visualize drug-protein complexes using build_viewer.py, PubChem, and OpenFold3 NIM. Use when asked to show a molecular structure, drug target, or protein visualization.	openclaw requires bins python3 curl

Molecular Visualization

Generate 3D protein-ligand visualizations using the build_viewer.py script. The script handles the full pipeline:

1. Drug SMILES -- looked up automatically from PubChem
2. Protein target -- resolved from a built-in drug-target table (or pass --sequence manually)
3. Structure prediction -- protein + drug sent to OpenFold3 NIM for co-structure prediction
4. 3D viewer -- self-contained HTML with jQuery + 3Dmol.js inlined, saved to canvas

Usage

Simplest form (target auto-resolved):

python /sandbox/clinical-intelligence/scripts/build_viewer.py --drug metformin

With explicit sequence (for drugs not in the built-in table):

python /sandbox/clinical-intelligence/scripts/build_viewer.py --drug drugname --sequence AMINOACIDSEQ --title "Custom Title"

Options

Flag	Required	Description
--drug	Yes	Drug name for PubChem SMILES lookup (e.g. metformin)
--sequence	No	Amino acid sequence of protein target. Auto-resolved if omitted.
--title	No	Custom viewer title
--output	No	Custom output path (defaults to ~/.openclaw/canvas/{drug}_complex.html)
--openfold-host	No	Override OpenFold3 host IP (defaults to 172.17.0.1)

Built-in drug targets

The script knows these drugs and auto-resolves their protein targets:

Drug	Target protein
metformin	Insulin B-chain
atorvastatin	HMG-CoA reductase
rosuvastatin	HMG-CoA reductase
lisinopril	ACE
enalapril	ACE
losartan	Angiotensin II receptor type 1
amlodipine	L-type calcium channel Cav1.2
empagliflozin	SGLT2
semaglutide	GLP-1 receptor

For any drug in this table, just pass --drug and the script does the rest.

Drugs NOT in the table

If the drug is not listed, the script exits with an error and prints the list of known drugs. In that case, you need to provide --sequence explicitly. Tell the user the drug is not in the built-in table and that you need a protein target sequence to proceed.

Drugs that cannot be visualized

Biologics, enzyme mixtures, or complex formulations that PubChem cannot resolve to a single SMILES (e.g. pancrelipase, insulin glargine) will still get protein-only structure prediction -- the script handles this gracefully by predicting without a ligand.

Output

The script saves an HTML viewer to canvas. Link it in your response as a markdown hyperlink:

[View 3D structure](http://localhost:18789/__openclaw__/canvas/metformin_complex.html)

Confidence Scores

The viewer header displays OpenFold3 scores:

• Confidence -- overall prediction confidence (higher = better)
• pLDDT -- per-residue local confidence (0-100, >70 is good)
• pTM -- predicted template modeling score (0-1)
• ipTM -- interface predicted TM-score (complexes only, measures protein-ligand interface quality)