Detect and rank druggable pockets on a protein structure

Point Claude Code at a coordinate file with no bound ligand; get back a ranked pocket table with geometry and learned-ligandability scores kept side by side, the lining residues of each pocket, and a docking box ready for the screening step.

   
Problem class Data analysis
Subject areas Integrative Structural and Computational Biology, Drug Repurposing and Discovery
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

You have a structure for a new target and no co-crystallized ligand to tell you where to aim. Every downstream structure-based step needs a site: docking needs a box, a virtual screen needs a receptor definition, a mutagenesis plan needs lining residues. Guessing from a cartoon render is how projects end up screening against a crystallization-additive cleft for three months.

The tooling exists and is mature — fpocket enumerates cavities geometrically, P2Rank scores surface points for learned ligandability — but the two answer different questions, install differently, and disagree often enough that reconciling them by hand is the actual work. Worse, the failure modes are silent: druggability models were trained on ligand-bound structures, so an apo or cryptic site gets a low score for reasons that have nothing to do with whether a drug can bind there. Solved looks like a committed script that produces one pockets.csv per structure — rank, volume, druggability score, lining residues, both tools’ verdicts as separate columns — plus a docking box you can hand straight to Vina, and a provenance record naming the tool versions and the input hash.

Rung 2 — one Claude Skill, Binding Site Detection, which drives fpocket, P2Rank and (optionally) the CASTp / DoGSiteScorer web services and knows how to read their disagreements. Keep the run local unless your data-sharing policy permits sending coordinates to an academic web server.

  1. Install the skill. Verbatim steps are on the catalog page (clone GPTomics/bioSkills, run ./install-claude.sh --categories "structural-biology" or copy structural-biology/binding-site-detection into ~/.claude/skills/). Install fpocket 4.1+ and P2Rank 2.4+ when prompted on first use — the skill will not silently substitute one for the other.

  2. Vet the receptor before you search it. Pockets are only as real as the atoms lining them. Run the Structure Validation skill on an experimental file, or — for an AlphaFold model — triage the model first: pocket-lining side-chain rotamers are among the least reliable atoms in a predicted structure, and a low-pLDDT region will produce cavities that do not exist. Strip waters, ions and crystallization additives with the Structure Preparation skill before detection, or fpocket will happily wrap a pocket around a PEG molecule.

  3. Enumerate, then rank — as two separate columns. Ask for both, and do not let either be collapsed into a single number:

    Structure: receptor_clean.pdb (apo, chain A only).

Using the binding-site-detection skill:
  1. Run fpocket. For each pocket report: rank, volume (A^3),
     number of alpha spheres, mean local hydrophobic density,
     fpocket druggability score, and the lining residue list
     (chain + resnum + resname).
  2. Run P2Rank on the same file. For each predicted pocket
     report: rank, P2Rank score, pocket center (x,y,z), and
     lining residues.
  3. Match the two sets by centroid proximity (<= 5 A) and
     by lining-residue Jaccard overlap. Emit ONE table with
     columns: site_id, fpocket_rank, fpocket_volume,
     fpocket_drugscore, p2rank_rank, p2rank_score,
     n_shared_residues, agreement (both | fpocket_only |
     p2rank_only).
  Do not average the two scores into a single ranking.

    Sites where both tools agree are the safe bets. fpocket_only sites are usually shallow grooves or additive clefts; p2rank_only sites are worth a look precisely because P2Rank scores surface chemistry rather than enclosure.

  4. Anchor against what is already known. Before trusting rank 1, check the target’s annotated sites — catalytic residues, cofactor sites, a homolog’s co-crystal ligand. Pull them with the UniProt MCP server (Binding site / Active site features) or find a holo structure with the PDB MCP server and compare its ligand centroid to your predicted centers. Add a known_site_overlap column. If the tools miss a site you know is real, that is the calibration you need before believing the novel ones.

  5. Emit the docking box. The point of the exercise:

    For site_id S1, write docking_box.json with:
  center: [x, y, z] (pocket centroid),
  size:   [sx, sy, sz] (pocket bounding box + 4 A padding
          on each axis, rounded to 0.5 A),
  lining_residues: [...],
  source: "fpocket+P2Rank consensus".

    That file drops straight into AutoDock Vina / smina, or seeds the pocket for a DiffDock run and the affinity-ranking recipe.

  6. Escalate to cryptic pockets only if the apo search comes up empty. If the surface looks flat and the target is nonetheless known to be ligandable, run mdpocket over a conformational ensemble — a trajectory from the GROMACS MCP server or OpenMM MCP server, set up via the MD-simulation recipe — and report pocket occurrence frequency across frames, not a single-frame score. This is the step that turns a laptop job into a GPU job; do not reach for it first.

  7. Commit the artifact. Version detect_pockets.sh (the fpocket and P2Rank invocations with their exact flags) and reconcile_pockets.py (the matching and table emission), alongside pockets.csv, docking_box.json, the prepared receptor_clean.pdb, and a provenance.json recording: fpocket and P2Rank versions, the P2Rank model name, the input PDB/AlphaFold accession and its file sha256, the preparation steps applied, whether any web service was called, and the model id that authored the script. Pocket rankings shift between tool versions; without the versions recorded you cannot tell a real conformational difference from a scorer update. See Reproducible, provenance-tracked AI analysis.

Why this assembly

Rung 2, and it stops there. The scientific content is two established command-line tools plus the judgement to read them against each other; the skill supplies the invocation details and the interpretation discipline (geometric cavity ≠ functional site; holo-trained scores under-detect apo pockets). Rung 1 — plain Claude Code — can shell out to fpocket if you have already installed and learned it, but will improvise flag choices and, worse, will confidently rank a crystallization-additive cleft first because nothing in its context says not to. Rung 3 is not warranted: the neighbouring components here (validation, preparation, docking) are separate recipes with their own outputs, not co-resident tools in one harness. Rung 4 is plainly excessive for a single-structure analysis whose output is one table.

Availability

Fully open. The skill is MIT-licensed; fpocket, P2Rank and mdpocket are open-source and run entirely on your machine. Two optional components are external academic web services — CASTp and DoGSiteScorer (via the ProteinsPlus server) — so an unpublished or confidential structure should stay on the local fpocket/P2Rank path. There is no local substitute in the catalog for DoGSiteScorer’s SVM druggability model; if you need a druggability number and cannot upload, use fpocket’s own druggability score and treat it as coarser. P2Rank requires a Java runtime.

Compute requirements

Laptop. P2Rank predicts on a single protein in under a second; fpocket is comparable, so a whole-PDB-entry scan is seconds and a few-hundred-structure batch runs in minutes on one core. Memory is negligible (tens of MB per structure). The one heavy path is step 6: mdpocket needs an MD trajectory first, which is a GPU job of hours to days depending on system size and the timescale of the cryptic opening — budget that separately, and note that the pocket analysis itself remains cheap once frames exist.

Evidence

Proposed. No documented attempt at this exact assembly — Claude Code driving the Binding Site Detection skill through an enumerate-then-rank-then-reconcile workflow — is known as of 2026-08-01. The components and the design of the workflow are well benchmarked:

  • Rank with a learned scorer, enumerate with geometry. In the largest binding-site benchmark to date — 13 predictors and 15 variants against the LIGYSIS dataset of ~30,000 protein–ligand complexes — fpocket predictions rescored by PRANK or DeepPocket gave the highest recall (60%), versus 39% for the weakest method tested; stronger scoring schemes alone improved recall by up to 14% and precision by up to 30% (Utgés & Barton, J. Cheminform. 2024). That result is why step 3 separates enumeration from ranking rather than trusting one tool’s ordering.
  • P2Rank as the ranking layer. P2Rank outperformed fpocket, SiteHound, MetaPocket 2.0 and DeepSite in its original evaluation, at under 1 s per protein (Krivák & Hoksza, J. Cheminform. 2018) — the speed is what makes batch use practical.
  • The apo caveat is measured, not folklore. CryptoBench assembles 1,107 apo–holo pairs with substantial binding-site conformational change and shows that holo-based evaluation “yields unrealistic performance expectations” for cryptic sites, with P2Rank explicitly not tailored to them (Škrhák et al., Bioinformatics 2024). This grounds step 6 and the warning against reading a low score as “not druggable”.
  • Membrane targets degrade. Across GPCR and ion-channel sets, every method tested scored worse on membrane-embedded interfaces than on a soluble PDBBind set (best-case normalized DCC / DVO 0.33–0.72) (Pliushcheuskaya & Künze, J. Chem. Inf. Model. 2025). If your target is a GPCR or channel, weight the known-site anchor in step 4 more heavily than the scores.

The missing link is a head-to-head of an agent-driven reconciliation against a structural biologist doing the same by hand.

Alternatives considered

PrankWeb (hosted). The same P2Rank engine behind a web server, now with an integrated AutoDock Vina docking module, UniProt-ID input that pulls the AlphaFold model for you, and ChimeraX/PyMOL export (Polák et al., Nucleic Acids Res. 2025; AlphaFold and Docker support landed in PrankWeb 3). Reach for it for a one-off look at a public target — it is faster than installing anything. Reach for this recipe when you need a batch, an audit trail, or the structure never to leave your machine.

Skip detection entirely. If a holo structure of your target or a close homolog exists, superpose it and take the ligand’s position as the site — cheaper and more reliable than any prediction. Superpose two protein structures is that path, and step 4 here is its degenerate case. Detection earns its keep only when no bound-state reference exists.

Blind docking. DiffDock and similar can search the whole surface without a box. That trades a defensible, inspectable site hypothesis for a pose that may be anywhere, and it does not give you lining residues to design mutants against. Use it as a cross-check on the sites this recipe ranks, not as a replacement.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.