DiffDock (Claude Skill)

Claude skill that drives DiffDock — a diffusion-generative model that predicts 3D protein–ligand binding poses directly from a protein structure (PDB) and a ligand SMILES, without exhaustive grid search.

   
Type Claude Skill
Supplier K-Dense Inc. (community OSS); DiffDock model from gcorso/DiffDock (MIT)
Availability GA — actively maintained in the K-Dense scientific-agent-skills collection
Pricing Free / OSS skill (MIT); DiffDock model weights MIT
Capabilities Read/Write — Claude executes DiffDock locally via Python/Bash; writes generated poses (SDF / PDB) and confidence CSVs to disk

How to install

  • Also packaged in the SciAgent-Skills collection (jaechang-hits (community OSS, CC BY 4.0)): clone jaechang-hits/SciAgent-Skills and run /plugin install sciagent-skills in Claude Code (or copy skills/structural-biology-drug-discovery/diffdock into ~/.claude/skills/).
  • Claude Code / Claude.ai — Skills CLI (recommended): 
    npx skills add K-Dense-AI/scientific-agent-skills

    Installs the K-Dense collection; enable the diffdock skill when prompted (also works in Cursor/Codex via the Agent Skills spec; requires Node ≥ 18).

  • Claude Code / Claude Desktop — manual clone: 
    git clone https://github.com/K-Dense-AI/scientific-agent-skills
cp -r scientific-agent-skills/skills/diffdock ~/.claude/skills/
git clone https://github.com/gcorso/DiffDock.git
cd DiffDock && conda env create --file environment.yml && conda activate diffdock

    Project-scoped alternative: copy into .claude/skills/ (inside your project) instead of ~/.claude/skills/. DiffDock itself ships as a separate conda environment; the skill’s setup_check.py validates Python, PyTorch + CUDA, PyTorch Geometric, RDKit, and ESM before any docking run.

What it does

SKILL.md plus assets (batch_template.csv, custom_inference_config.yaml), references (confidence_and_limitations.md, parameters_reference.md, workflows_examples.md), and scripts (analyze_results.py, prepare_batch_csv.py, setup_check.py) with recipes for:

  • Pose generation — take a protein PDB and a ligand SMILES, generate N diffusion-sampled binding poses (default 10–40) with per-pose confidence scores
  • Virtual screening — batch many ligands against one protein via prepare_batch_csv.py; rank by confidence
  • Pose analysis — analyze_results.py filters, clusters, and exports top-K poses as SDF / PDB for downstream MM/PBSA or visualization
  • Workflow handoff — confidence-filtered poses feed molecular-dynamics for refinement, deepchem / medchem for re-scoring, and molecule-mcp (PyMOL / ChimeraX) for visualization

Primary use cases: blind docking against AlphaFold-predicted targets, lead-optimization pose prediction, virtual screening pre-filter ahead of MM/PBSA or FEP, allosteric-site exploration when binding pockets are uncertain.

Notes

  • Poses only, not affinity. DiffDock predicts binding geometry and a confidence score, not binding affinity (ΔG, K_d). Pair with GNINA / Vina / MM-GBSA / FEP for affinity ranking — the SKILL.md is explicit about this.
  • GPU strongly recommended. Diffusion sampling on CPU is slow; the upstream environment ships with CUDA-pinned PyTorch. The K-Dense molecular-docking workflow example chains DiffDock with DeepChem rescoring and MedChem filtering.
  • Confidence interpretation. Per confidence_and_limitations.md, confidence > 0 typically indicates a high-quality pose; confidence < -1.5 is unreliable. Always inspect top-K visually.
  • Pairs with the molecular-dynamics, molecule-mcp, alphafold, pdb, and deepchem catalog entries. The skill is documentation plus Python recipes — Claude drives DiffDock locally via Bash/Python; the skill does not ship binaries.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.