Diffdock Molecular Docking

Source: https://github.com/aipoch/medical-research-skills

DiffDock Molecular Docking

When to Use

• Blind docking when you have a protein structure (PDB) and a ligand (SMILES) but no known binding site.
• Pose prediction to generate multiple plausible 3D binding conformations and rank them.
• Virtual screening support to quickly evaluate candidate ligands by predicted binding poses and confidence.
• Drug discovery workflows where you need automated docking outputs (SDF poses + scores) for downstream analysis.
• Batch/advanced docking when running many ligand–protein pairs or using alternative inputs (e.g., sequence-based workflows; see references/workflows_examples.md).

Key Features

• Diffusion generative sampling to produce diverse ligand binding poses.
• Confidence model scoring to rank predicted poses.
• Simple CLI inference for single protein–ligand docking.
• Batch/advanced workflows documented in references/workflows_examples.md.
• Structured outputs including ranked SDF pose files and a confidence score report.

Dependencies

• Python (version not specified)
• PyTorch (version not specified)
• PyTorch Geometric / PyG (version not specified)
• RDKit (version not specified)
• ESM (version not specified)

Example Usage

1) Verify the Environment

python scripts/setup_check.py

2) Run Standard Inference (Single Docking)

Dock a single ligand (SMILES) to a protein structure (PDB) and write results to an output directory:

python scripts/inference_runner.py \
  --protein ./data/protein.pdb \
  --ligand "CC(=O)Oc1ccccc1C(=O)O" \
  --out_dir ./results

Arguments

• --protein: Path to the protein PDB file.
• --ligand: Ligand SMILES string.
• --out_dir: Output directory (default: results/).

3) Outputs

After inference, the tool produces:

• Ranked SDF pose files (e.g., rank1.sdf, rank2.sdf, ...), each containing a predicted 3D binding pose.
• Confidence score report: confidence_scores.txt, listing the score for each ranked pose.

Implementation Details

• Pose generation: Uses a diffusion-based generative model to sample multiple candidate ligand poses relative to the protein target.
• Ranking: A separate confidence model assigns a score to each sampled pose; poses are sorted by this score and saved as rank*.sdf.
• Parameterization:
  • For the complete CLI argument list and defaults, see references/parameters_reference.md.
  • For confidence interpretation, known limitations, and expected accuracy/scope, see references/confidence_and_limitations.md.
• Advanced workflows: Batch processing and alternative input configurations are documented in references/workflows_examples.md.