Triage an AlphaFold model for structure-based drug design

Hand Claude Code a UniProt accession; get back the AlphaFold predicted structure, a pLDDT map, the confidence verdict on the binding pocket (or any user-named region), and a one-line recommendation on whether the model is safe to use as a docking template — or whether to fall back to an experimental PDB structure or to refine first.


Problem class	Knowledge synthesis
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

AlphaFold has solved “do I have a structure” for ~200 million proteins. It has not solved “is this structure fit for my modelling question.” Docking against a high-pLDDT loop is fine; docking into a pocket that AlphaFold modelled at pLDDT 55 is wishful thinking. The standard pre-flight checks — global pLDDT mean and distribution, per-residue confidence over the binding pocket, presence of disordered tails, comparison to any available experimental PDB structure — are quick to run individually but tedious to assemble for every new target. A computational chemist preparing a virtual screen, an antibody engineer picking a docking input, or a structural biologist scoping a refinement project all need the same pre-flight in 30 seconds, not 30 minutes of clicks across the EBI portal, RCSB, and PyMOL.

Solved looks like: one prompt with a UniProt accession (and optionally a pocket-residue list); one structured triage card out — global pLDDT stats, region-level pLDDT, low-confidence stretches flagged, the closest experimental PDB entry named, and a go / refine / fall-back-to-PDB verdict.

Recommended approach

This recipe is rung 2 — one MCP server, the AlphaFold MCP Server, wraps the ~25 EBI AlphaFold API endpoints Claude needs. The PDB cross-check is optional and uses the companion PDB MCP Server when available.

Install the AlphaFold MCP server. Verbatim install commands live on the catalog page; the short version is a git clone, npm install && npm run build, then claude mcp add --transport stdio alphafold-server -- node /path/to/build/index.js. The server uses the public EBI API — no auth, no quota.

Pull the structure and confidence summary. A minimal prompt:

Target: UniProt P38398 (BRCA1).

Using the alphafold-server MCP:
  1. get_structure for P38398; capture length, model
     identifier, and modelling date.
  2. get_confidence_scores for P38398. Print:
       mean pLDDT, median pLDDT, % residues >90, % 70–90,
       % 50–70, % <50.
  3. analyze_confidence_regions to enumerate contiguous
     stretches of residues with pLDDT < 70, with start
     and end residue numbers.

Emit a markdown table of the regions and a one-paragraph
global verdict ("model X% high-confidence; N disordered
stretches comprising Y residues").

Score the pocket (or any region of interest). If the user has a pocket residue list — from a homologous PDB structure, a paper, or a pocket-prediction tool — score it explicitly:

Pocket residues for BRCA1 BRCT domain (paste-or-derive):
  1646, 1648, 1655, 1660, 1701, 1704, 1731, 1736, 1760.

For each pocket residue, report its pLDDT from the
get_confidence_scores payload. Compute pocket-pLDDT
mean and minimum. Apply the rule of thumb:
  - pocket mean ≥ 85 and min ≥ 70  →  docking-ready
  - pocket mean 70–85, min 50–70   →  usable with caveats; flag
    low-confidence side chains for refinement
  - pocket mean < 70 or min < 50   →  not docking-ready;
    fall back to experimental PDB or co-folding model
Print the verdict.

(Optional) Cross-check with experimental PDB structures. If you have the PDB MCP server registered, add:

Using pdb-server, run search_by_uniprot for P38398.
Return the top 5 PDB entries by resolution. For each:
  PDB ID, resolution (Å), chain coverage of the AlphaFold
  residue range, deposit year, ligand-bound or apo.
If any PDB entry covers the pocket residues above at
resolution ≤ 2.5 Å, recommend it over the AlphaFold model
as the primary docking template.

Export for a downstream tool. AlphaFold MCP can prepare PyMOL or ChimeraX export commands directly:

Use export_for_pymol on P38398 to produce a PyMOL session
script that colours the cartoon by pLDDT (blue >90,
cyan 70–90, yellow 50–70, orange <50) and zooms on the
pocket residue list. Save the script as
structures/P38398_triage.pml.

Swap export_for_pymol for export_for_chimerax if that’s the visualization stack of choice.

Persist the triage card. Ask Claude Code to write structures/<UniProt>_triage_<date>.md with the global stats, low-pLDDT region table, pocket verdict, PDB alternatives, and the PyMOL/ChimeraX script path.

Why this assembly

Rung 2 of the simplicity ladder. The AlphaFold MCP Server wraps every API call this triage needs — structure retrieval, pLDDT scoring, region analysis, visualization export — behind a single MCP. Rung 1 (plain Claude Code) would hit the same EBI endpoints with raw HTTP but loses the typed tool surface and the built-in pLDDT region analysis. Rung 3 (an extra toolbelt) is only justified when the optional PDB cross-check is in scope; that’s one extra tool and remains a thin escalation, not a harness. No rung-4 autonomous-science system is required — the workflow is shallow, the output is a card, and per-claim provenance to AlphaFold/PDB accessions is the value.

Availability

Fully open. The AlphaFold MCP server is MIT-licensed; EBI’s AlphaFold Protein Structure Database is free for academic use (commercial users should consult the EBI terms). The PDB MCP server uses the public RCSB REST API. No subscription required. Local Node and npm are the only environment dependencies.

Compute requirements

Laptop. A single triage card returns in under 30 seconds: structure retrieval is a few hundred kilobytes; pLDDT analysis is in-process; PyMOL/ChimeraX export is a text script. Running PyMOL or ChimeraX to render the session — if you do — adds whatever those clients cost on your machine.

Evidence

Proposed. No documented end-to-end LLM-orchestrated triage workflow using the AlphaFold MCP server in peer-reviewed literature is known as of 2026-06-09. The component pieces are well-validated:

AlphaFold MCP Server — the server’s pLDDT-analysis tools wrap the same EBI API that underpins the published AlphaFold Protein Structure Database (Varadi et al., Nucleic Acids Res. 2022, 50:D439; 2024 update 50:D368).
pLDDT-thresholding rules of thumb — the docking-readiness cutoffs above follow the practical benchmarks established for AlphaFold-Multimer and protein-protein docking (Bryant et al., Nature Communications 2022, 13:1265; follow-up assessments through 2024–2025 reaffirm interface-pLDDT ≥85 as the strong-confidence regime, 70–85 as the caveat regime).
AlphaFold for small-molecule docking — community benchmarks repeatedly find that high global pLDDT does not guarantee virtual-screen success; pocket-residue pLDDT and side-chain rotamer fidelity matter more (Karelina et al., J. Chem. Inf. Model. 2023, 63:6219, and 2024–2025 follow-ups). The recipe encodes that distinction in step 3.
AlphaFold MCP usage — referenced in several MCP-server directories and a Skywork.ai walkthrough; not yet cited in peer-reviewed work.

A peer-reviewed benchmark of “Claude + AlphaFold MCP triage” vs a hand-built notebook is the missing link. The component-level evidence behind every claim — the API, the pLDDT cutoffs, the rule-of-thumb thresholds — is well-established.

Alternatives considered

Rung 1 — plain Claude Code + EBI HTTP. Possible but throws away the typed MCP tool surface and the analyze_confidence_regions tool. Pick this only if your environment forbids MCP servers.
Local AlphaFold inference. Running AlphaFold 2/3 locally to predict a new structure is the right escalation when EBI’s pre-computed entry is for the wrong species, isoform, or splice variant. That requires GPU compute and is outside this recipe’s scope.
Co-folding (AlphaFold-Multimer, Boltz-2, RoseTTAFold All-Atom). Better than monomer AlphaFold for ligand-bound or complex states. Not yet wrapped as a Claude-installable component in catalog/tools/; see Missing components in the curator state.
Skip AlphaFold, use the PDB structure directly. When a high-resolution apo or holo structure covering the pocket exists, prefer it. Step 4 of the recipe is exactly that check.

Sources

Augmented-Nature/AlphaFold-MCP-Server — verified 2026-05-27 (this run).
Varadi M. et al., “AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models,” Nucleic Acids Res. 2022 — published 2021-11.
Varadi M. et al., “AlphaFold Protein Structure Database in 2024,” Nucleic Acids Res. 2024 — published 2023-11.
Bryant P. et al., “Improved prediction of protein-protein interactions using AlphaFold2,” Nature Communications 2022 — published 2022-03.
Karelina M. et al., “How accurately can one predict drug binding modes using AlphaFold models?” J. Chem. Inf. Model. 2023 — published 2023-09.
EBI AlphaFold Protein Structure Database — verified 2026-05-27 (this run).

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.