Use DiffDock Online
Commercially Available DiffDock No-Code Web Server
DiffDock: Diffusion-Based Blind Molecular Docking for Protein-Ligand Pose Prediction
DiffDock is a diffusion generative model for molecular docking that predicts how a small-molecule ligand binds to a protein. It was introduced by Gabriele Corso, Hannes Stärk, Bowen Jing, Regina Barzilay and Tommi Jaakkola of MIT CSAIL and presented at ICLR 2023. Instead of searching for a single best pose, DiffDock samples many candidate poses and ranks them with a learned confidence model.
Molecular docking is central to computational drug discovery, yet it is hard. Traditional search-based docking tools score a large number of candidate poses, which is slow, and earlier deep learning approaches that regress a single pose have struggled with accuracy. The authors argue that docking is better framed as a generative modeling problem, because the task is inherently uncertain and several poses can be plausible.
Fact | Detail |
|---|---|
Task | Blind docking of a small molecule to a protein, with no binding pocket specified |
Model type | Diffusion model over ligand poses plus a separate confidence model |
Inputs | A protein structure and a ligand with a seed conformation (typically generated with RDKit) |
Top-1 success (RMSD below 2 Å) on PDBBind | 38.2% with 40 samples |
Venue | ICLR 2023 |
How DiffDock Works: Diffusion Over Ligand Poses and Confidence Ranking
A ligand pose is described by its position, its orientation and its torsion angles. DiffDock maps the non-Euclidean space of ligand poses onto a product space of translations, rotations and torsional degrees of freedom, and runs an efficient diffusion process over that space.
Sample random poses: the ligand starts from random positions, orientations and conformations around the protein.
Denoise with the score model: the diffusion model iteratively moves the ligand toward plausible binding poses.
Rank with the confidence model: a separate model, trained with cross-entropy to predict whether a pose lies within 2 Å of the true one, scores each sampled pose.
Select the top poses: the highest-confidence poses are returned for inspection.
DiffDock Results: Docking Accuracy, Speed and Confidence Estimates
On the PDBBind benchmark, DiffDock reaches a top-1 success rate of 38%, compared with 23% for traditional docking methods and 20% for earlier deep learning methods, as reported in the paper.
Setting | Top-1 success (RMSD below 2 Å) | Top-5 success |
|---|---|---|
40 samples | 38.2% | 44.7% |
10 samples | 35.0% | 40.7% |
Confidence and Runtime
Selective accuracy: on the one third of complexes where the confidence model is most confident, success rises to 83%.
Speed: about 10 seconds for 10 samples and 40 seconds for 40 samples per complex on average, which the paper reports as 3 to 12 times faster than GNINA on GPU.
Computationally folded proteins: without retraining, DiffDock is reported to reach 21.7% precision on predicted structures, against a maximum of 10.4% for previous methods.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with DiffDock on Tamarind Bio
Blind docking of new ligands: place a small molecule on a protein when the binding pocket is unknown.
Hit prioritization: use confidence scores to decide which predicted complexes deserve follow-up.
Docking to predicted structures: the paper shows the model can be applied to computationally folded proteins.
Pose hypotheses for design: generate several candidate binding modes to inform ligand optimization.
How to Use DiffDock on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and open DiffDock.
Provide the protein: upload a protein structure file, or use a predicted structure.
Provide the ligand: enter or upload the small molecule, for example as a SMILES string, if exposed.
Set parameters: choose the number of poses to sample, if exposed. The paper compares 10 and 40 samples.
Run the job: submit the run and wait for it to finish.
Download outputs: retrieve the ranked ligand poses and their confidence scores.
Interpret results: favor poses with higher confidence, and treat low-confidence complexes with caution.
Use downstream: take the top poses into rescoring, visual inspection or further design.
Things to Keep in Mind
The paper assumes access to the holo (bound) protein structure, although the score model uses only alpha-carbon positions.
Protein flexibility is not modeled, and side-chain flexibility is left to future work.
DiffDock is slower than one-shot regression methods such as EquiBind.
Success rates on folded structures are lower than on crystal structures.
Source: Corso G, Stärk H, Jing B, Barzilay R, Jaakkola T. DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking. ICLR 2023. arXiv:2210.01776 (https://arxiv.org/abs/2210.01776). Code: https://github.com/gcorso/DiffDock.