Diffdock

DiffDock: Diffusion-Based Molecular Docking

Overview

DiffDock predicts 3D protein-ligand binding poses using a diffusion model. Given a protein structure and a ligand SMILES string, it generates multiple candidate poses ranked by a confidence score.

Key properties:

• Blind docking — no binding site specification required; searches entire protein surface
• Diffusion model — generates diverse poses via iterative denoising (score-based diffusion)
• SE(3)-equivariant — architecture respects 3D symmetry (rotations, translations)
• Confidence model — separate network ranks poses by predicted accuracy (not binding affinity)
• Current default: DiffDock-L (ICLR 2024) — improved generalization over original (ICLR 2023)

Important distinction: The confidence score predicts pose quality (RMSD to true binding pose), not binding affinity. High confidence ≠ strong binder.

Installation

DiffDock requires a conda environment — no pip-only install available:

git clone https://github.com/gcorso/DiffDock.git
cd DiffDock
conda env create --file environment.yml
conda activate diffdock

Or with Docker (includes GPU support):

docker pull rbgcsail/diffdock
docker run -it --gpus all --entrypoint /bin/bash rbgcsail/diffdock
micromamba activate diffdock

Key dependencies: PyTorch 1.13 + CUDA 11.7, torch-geometric 2.2, e3nn 0.5.1, RDKit, fair-esm (ESM-2 for protein embeddings), pytorch-lightning.

First run: precomputes SO(2)/SO(3) distribution caches (~2 min). Not repeated.

Core Usage

Single complex

cd DiffDock
python -m inference \
  --config default_inference_args.yaml \
  --protein_path protein.pdb \
  --ligand_description "COc(cc1)ccc1C#N" \
  --out_dir results/my_docking

From protein sequence (ESMFold folds it automatically)

python -m inference \
  --config default_inference_args.yaml \
  --protein_sequence "MVHLTPEEKSAVTALWGKVNVDEVGGEALGRLLVVYPWTQRFFESFGDLSTPDAVMGNPKVKAHGKKVLGAFSDGLAHLDNLKGTFATLSELHCDKLHVDPENFRLLGNVLVCVLAHHFGKEFTPPVQAAYQKVVAGVANALAHKYH" \
  --ligand_description "CC(=O)Nc1ccc(O)cc1" \
  --out_dir results/my_docking

Batch processing (CSV input)

python -m inference \
  --config default_inference_args.yaml \
  --protein_ligand_csv complexes.csv \
  --out_dir results/batch

CSV format (complexes.csv):

complex_name,protein_path,ligand_description,protein_sequence
complex_A,proteins/target_A.pdb,CC(=O)Nc1ccc(O)cc1,
complex_B,proteins/target_B.pdb,data/ligand_B.sdf,
complex_C,,COc1ccc(C#N)cc1,MVHLTPEEKSAVTALWG...

Ligand from SDF file

python -m inference \
  --config default_inference_args.yaml \
  --protein_path protein.pdb \
  --ligand_description ligand.sdf \
  --out_dir results/my_docking

Key Parameters

Parameter	Default	Description
--samples_per_complex	10	Poses to generate per complex
--inference_steps	20	Denoising steps (fewer = faster, less refined)
--batch_size	10	Parallel complexes (reduce if OOM)
--no_final_step_noise	True	No noise in final step (better quality)
--save_visualisation	False	Save reverse-diffusion trajectory PDB

Speed vs quality:

Mode	--inference_steps	--samples_per_complex	--batch_size
Fast screening	5	5	20
Default	20	10	10
High quality	40	20	5

GPU: ~2–5 sec/pose. CPU: ~30–60 sec/pose.

Output Format

results/my_docking/
└── complex_0/
    ├── rank1_confidence0.43.sdf     ← best pose, high confidence
    ├── rank2_confidence-0.12.sdf    ← 2nd best, moderate
    ├── rank3_confidence-1.23.sdf    ← 3rd, low confidence
    └── ...                          ← up to samples_per_complex

• Files are SDF format — open directly in PyMOL, ChimeraX, or RDKit
• Ranked by confidence score (rank1 = highest)
• Confidence score is in the filename: rank{N}_confidence{score}.sdf

Confidence Score Interpretation

Score	Confidence	Meaning
> 0	High	Model is confident the pose is accurate
-1.5 to 0	Moderate	Plausible but uncertain
< -1.5	Low	Unreliable; treat with caution

Caveats:

• Score predicts RMSD accuracy, not binding affinity
• Thresholds shift downward for large ligands, apo proteins, or unusual binding sites
• Always inspect top poses visually — confidence is a heuristic, not ground truth
• For affinity estimation, follow up with MM-PBSA, GNINA, or FEP

Practical Notes

• No pocket required — DiffDock docks to the entire protein surface
• One ligand at a time — run separately for each ligand; does not support co-docking
• Rigid protein — assumes fixed backbone; for flexibility, relax the pose with MD
• Clean PDB — remove other ligands and waters from the binding site before docking
• OOM — reduce --batch_size to 1 or 5 for large proteins (>1000 residues)
• Validate setup — first run on a known PDB complex with its co-crystal ligand; check RMSD

Scripts

• scripts/dock.py — wrapper that runs DiffDock and parses results; see scripts/dock.py --help

Resources

• GitHub: https://github.com/gcorso/DiffDock
• DiffDock (original): Corso et al., ICLR 2023 — https://arxiv.org/abs/2210.01776
• DiffDock-L: Corso et al., ICLR 2024 — https://arxiv.org/abs/2402.18396

References

• references/cli-reference.md — full CLI parameter reference, output format details, confidence model, batch CSV format, DiffDock-L vs original