diffdock
Diffusion-based molecular docking. Predict protein-ligand binding poses from PDB/SMILES, confidence scores, virtual screening, for structure-based drug design. Not for affinity prediction.
Author
Category
Other ToolsInstall
Download and extract to your skills directory
Copy command and send to OpenClaw for auto-install:
DiffDock: A Molecular Docking Tool Based on Diffusion Models
Capability Overview
DiffDock is a molecular docking tool that uses diffusion models and deep learning to predict the three-dimensional binding poses of small-molecule ligands with protein targets. It is widely used in structure-based drug discovery and computational chemistry research.
Applicable Scenarios
When you need to predict how candidate compounds bind to a target protein for a drug development project, DiffDock can quickly generate 3D ligand binding poses and provide confidence scores to help select the most promising molecules. It supports PDB structure files or amino acid sequences as protein input, and SMILES strings or structure files as ligand input.
When screening hundreds to thousands of compounds, DiffDock supports batch processing mode and can efficiently dock an entire compound library. For large-scale screening tasks, protein embedding vectors can be precomputed to accelerate the workflow.
When researchers need to understand small-molecule–protein interaction mechanisms, DiffDock can generate diverse binding poses for structural analysis. Although it does not directly predict binding affinity, the generated poses can be used with scoring functions such as GNINA, MM/GBSA, or free energy calculation tools for downstream analysis.
Key Features
Using diffusion models and deep learning techniques, DiffDock can predict the 3D position and orientation of small-molecule ligands in protein binding pockets. It supports single-chain and multi-chain protein structures and can handle rigid and flexible ligands. Each docking run produces multiple candidate poses ranked by confidence.
Protein input supports PDB files or amino acid sequences (automatically folded via ESMFold). Ligand input supports formats such as SMILES, SDF, and MOL2. Outputs include ranked SDF structure files and detailed confidence scores for downstream analysis and visualization.
Provides a complete bulk docking workflow, including CSV-format batch input preparation, result confidence analysis scripts, and statistical summary export functions. Supports customizing sampling density, inference steps, temperature parameters, etc., allowing adjustments to prediction accuracy and diversity based on specific needs.
Frequently Asked Questions
Can DiffDock predict binding affinity (Kd, IC50)?
No. DiffDock predicts ligand binding poses (3D structures) and the model’s confidence in those predictions, but it does not directly predict binding affinity. To assess binding strength, it is recommended to score DiffDock-generated poses with GNINA, MM/GBSA, or free energy calculation tools.
Is DiffDock suitable for protein–protein docking?
No. DiffDock is designed for small-molecule ligands (typically 100–1000 Da) docking to proteins and is not suitable for protein–protein, protein–nucleic acid, or other large-molecule docking. For protein–protein docking needs, consider tools such as DiffDock-PP or AlphaFold-Multimer.
How can I improve the confidence of docking results?
Confidence is influenced by multiple factors: high molecular weight ligands (>500 Da) typically yield lower confidence; multi-chain proteins or novel protein families may affect prediction quality. It is recommended to increase the number of samples (from the default 10 to 20–40), try ensemble docking (using multiple protein conformations), ensure the protein structure is complete with no missing residues, and check whether the top 3–5 predictions show consensus.