protein-inspect

protein-inspect

Standardized inspection pipeline for protein structures. Combines deterministic feature extraction (gemmi, biotite, DSSP), a codified decision tree of structural-biology heuristics, and a fixed PyMOL view battery. Output is summary.yaml plus optional rendered images.

When to invoke

• User mentions a PDB ID (e.g. "1ubq", "look at 6vxx")
• User passes a structure file path (.pdb, .cif, .bcif)
• User asks "what does this protein look like / do / bind"
• User wants triage across multiple structures (RFdiffusion outputs, AF3 predictions)
• User asks about disulfides, oligomeric state, binding sites, or motifs in a structure

Usage

# basic — fetch from PDB and emit summary.yaml only
protein-inspect 1mbn

# with rendered view battery (slower, requires running PyMOL with claudemol plugin)
protein-inspect 1mbn --render-views

# with vision-judged render quality — scores each PNG against a per-view rubric
# and re-renders failing views with pre-declared knobs (best-of-N retained).
# Adds ~$0.05-0.25 per run on top of the render; requires ANTHROPIC_API_KEY or
# ANTHROPIC_AUTH_TOKEN. Per-view scores land in summary.yaml#visual[].judge.
protein-inspect 1mbn --render-views --judge-views

# pick a cheaper or stronger judge model (default: claude-sonnet-4-6)
protein-inspect 1mbn --render-views --judge-views --judge-model claude-haiku-4-5

# local file
protein-inspect /path/to/design.pdb --render-views

# with named motif for catalysis-focused analysis
protein-inspect 1kxj --motif His153,Asp166,Ser142 --render-views

# output directory
protein-inspect 1mbn --out ./analysis/1mbn/

Output layout

<out_dir>/
├── summary.yaml          # the semantic layer Claude reads
├── 1mbn.bcif             # coordinates, externalized
└── views/                # only when --render-views
    ├── 01_top.png
    ├── 02_side.png
    ├── 03_bfactor.png
    ├── 04_surface.png
    ├── 05_pocket_<lig>_<chain>.png
    └── 06_vicinal_ss_<chain>_<a>-<b>.png  # if vicinal disulfide present

How Claude should read the output

After invoking, read prompts/analyze.md and follow it. The prompt walks through the YAML and images in a fixed order so that analyses are reproducible across calls.

When to add --judge-views

Default: leave it off. The deterministic render battery is good enough for routine triage. Add --judge-views when:

• The user asks for "publication-quality" / "high-quality" / "clean" figures.
• The task hinges on a specific image being readable (e.g. "identify the residues coordinating this metal" — if the metal closeup is illegible, the analysis is blocked).
• You're triaging an unfamiliar design and want a sanity-check that the default views actually showed the relevant features.
• A first pass without --judge-views produced an image where you couldn't read what you needed.

Do NOT add --judge-views for fast sequence-style questions, batch screens across many structures, or anything where the YAML alone answers the question.

Requirements

• PyMOL running with the claudemol socket plugin (port 9880) — only needed for --render-views
• Python 3.12+ with deps from pyproject.toml

What this skill does NOT do

• Compute binding affinities (use FEP / MM-GBSA tools)
• Predict structure (use AlphaFold / Boltz / Chai)
• Detect novel spatial features outside the precomputed view battery
• Replace the deposition paper for mechanistic detail (it gives you the structural facts; biological interpretation still comes from literature or an MCP-backed annotation database)