DiffDock

Overview

DiffDock is a research implementation for predicting the three-dimensional structure of a small molecule bound to a protein. It supports the pose-generation stage of a molecular docking workflow rather than directly estimating binding affinity. The project README states that the repository runs DiffDock-L by default and directs users seeking the original DiffDock model to the repository history.

Inputs can combine a protein structure in .pdb format with a ligand represented as a SMILES string or an RDKit-readable file, such as .sdf or .mol2. Alternatively, users can provide a protein sequence, which the workflow folds with ESMFold. Individual complexes can be submitted directly, while a CSV describes multiple protein–ligand pairs for batch inference. Outputs include predicted binding structures and confidence scores. The repository also documents a local graphical interface for single complexes and links to a hosted Hugging Face Spaces interface.

Confidence scores describe the model’s confidence in a predicted pose, not binding affinity, and the README cautions against straightforward comparisons across complexes or protein conformations. Its interpretation guidance assumes inputs resembling the training setting: drug-like ligands and medium-sized proteins in near-bound conformations. The model was designed for small-molecule docking to proteins, not interactions between larger biomolecules. For downstream affinity estimation, the authors suggest relaxing predicted structures before applying separate scoring or free-energy methods. Local inference with supplied .pdb structures supports CPU execution, although the README recommends a GPU.

Key Features

• Predicts protein–small-molecule complex structures and provides pose confidence scores; the current repository defaults to DiffDock-L.
• Accepts protein `.pdb` files or protein sequences folded through ESMFold.
• Accepts ligand SMILES strings and RDKit-readable molecular files, including `.sdf` and `.mol2`.
• Supports single-complex inference and batch inference from CSV records containing protein and ligand inputs.
• Provides a local single-complex graphical interface, a hosted Hugging Face Spaces interface, and documented Conda and Docker setup routes.
• Documents benchmark evaluation workflows for PDBBind, DockGen, and PoseBusters, including preparation of cached ESM2 embeddings.

Use Cases

• Intended evaluation: generate candidate binding poses for a drug-like ligand and a protein structure, then inspect the poses and their confidence scores.
• Intended evaluation: process a collection of protein–ligand pairs through CSV-based inference to assess suitability for a research docking workflow.
• Intended evaluation: explore sequence-based docking through ESMFold-generated protein structures while accounting for the README’s cautions about unbound conformations.
• Intended evaluation: use predicted structures as starting inputs for separate affinity-scoring or free-energy workflows, with relaxation as recommended by the authors.

How to Use

1. Read the official repository documentation, especially the inference and FAQ sections. Decide whether you need the default DiffDock-L model or the original model described through repository history.
2. For an initial single-complex trial, use the linked Hugging Face Spaces interface. For local work, follow the README’s Conda or Docker setup instructions; it also documents a local graphical interface.
3. Prepare a protein .pdb file or a sequence for ESMFold, plus a ligand SMILES string or RDKit-readable file. For multiple complexes, follow the documented CSV fields and consult data/protein_ligand_example.csv.
4. Follow the repository’s inference example to select the configuration, inputs, and output directory. Account for initial distribution-table caching; the README recommends GPU execution where available.
5. Inspect predicted structures and confidence scores using the FAQ’s qualifications. For a proposed downstream affinity workflow, evaluate relaxation and separate scoring methods rather than treating confidence as affinity. Consult the repository’s dataset and replication sections for benchmark-oriented evaluation.