Use EquiDock Online

Commercially Available EquiDock No-Code Web Server

EquiDock: Fast SE(3)-Equivariant Rigid Protein-Protein Docking with Deep Learning

EquiDock is an end-to-end deep learning model for rigid-body protein-protein docking: given two unbound protein structures, it predicts the 3D structure of their complex in a single forward pass. It was developed by Octavian-Eugen Ganea, Xinyuan Huang, Charlotte Bunne, Yatao Bian, Regina Barzilay, Tommi Jaakkola and Andreas Krause, with the authors listed at MIT, ETH Zurich and Tencent AI Lab.

Protein-protein docking asks how two proteins arrange themselves when they bind, assuming each keeps its unbound shape. Classical docking tools typically sample and score a very large number of candidate poses, which makes them slow. EquiDock replaces this search with a learned model that directly outputs the rotation and translation that places one protein onto the other.

The authors designed the model so that its predicted complex does not depend on the starting positions or orientations of the two inputs, nor on which protein is treated as receptor or ligand. This is achieved through SE(3)-equivariance and commutativity constraints that the paper formalizes.

Property

Detail (from the paper)

Task

Rigid-body protein-protein docking

Inputs

Two unbound protein structures, represented as residue-level point clouds (alpha-carbon coordinates) with SE(3)-invariant features

Output

A rotation and a translation placing the ligand onto the fixed receptor

Core architecture

Independent SE(3)-equivariant graph matching network with keypoint matching and a differentiable Kabsch superimposition

Reported runtime

About 5 seconds on average per complex (local run), reported as an 80-500x speedup over baselines

Training data

DIPS (pretraining) and DB5.5 (fine-tuning and evaluation)

How EquiDock Predicts Protein Complexes

  1. Graph matching network: each protein is processed by a pairwise-independent SE(3)-equivariant graph network that exchanges messages within and between the two proteins and updates coordinates equivariantly.

  2. Keypoint selection: multi-head attention picks a set of keypoints on each protein. An optimal transport loss encourages them to line up with the midpoints of the binding pockets.

  3. Rigid placement: a differentiable Kabsch algorithm (SVD-based) converts the matched keypoints into the final rotation and translation.

  4. Training losses: the model combines a coordinate error on the ligand, the pocket-matching loss and a soft penalty against steric intersections.

The paper also introduces a surface feature that approximates residue depth without needing precomputed surface meshes.

EquiDock Benchmark Results and Speed

The authors compare EquiDock with ATTRACT, HDock, ClusPro and PatchDock on DIPS and DB5.5 using complex and interface RMSD (in angstroms). The table shows median values from the paper.

Method

DIPS C-RMSD

DIPS I-RMSD

DB5.5 C-RMSD

DB5.5 I-RMSD

ATTRACT

17.17

12.41

9.55

7.48

HDock

6.23

3.90

0.30

0.24

ClusPro

15.76

12.54

3.38

2.31

PatchDock

15.24

11.44

18.26

18.88

EquiDock

13.29

10.18

14.13

11.97

EquiDock is competitive rather than uniformly best: on DIPS it has lower medians than ATTRACT, ClusPro and PatchDock but not HDock, and on DB5.5 HDock and ClusPro are clearly more accurate. Its main advantage is speed, roughly 5 seconds per complex versus hundreds to thousands of seconds for the baselines. The authors also note that some baselines may have seen parts of the test sets, so the comparison may favor them.

Ablation findings

On DIPS, the full model reaches median C-RMSD / I-RMSD of 13.29 / 10.18. Removing the pocket loss worsens this to 15.91 / 12.01, removing the intersection loss to 15.49 / 11.09, and removing the surface features to 13.73 / 10.65.

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 EquiDock on Tamarind Bio

  • Rapid first-pass docking: generate a quick rigid-body placement of two unbound proteins before investing in slower, higher-accuracy methods.

  • Screening many pairs: because a prediction takes seconds, many protein pairs can be triaged for plausible binding arrangements.

  • Starting poses for refinement: use the predicted complex as an initial structure for downstream refinement or re-ranking, which EquiDock itself does not perform.

How to Use EquiDock on Tamarind Bio

  1. Open the tool: log in to tamarind.bio and open the EquiDock tool page.

  2. Provide the two proteins: supply the unbound receptor and ligand structures, as the model expects two protein structures (if exposed, as PDB files).

  3. Check the inputs: remember the model works on residue-level coordinates and assumes rigid bodies, so use unbound structures that are reasonably close to their bound shape.

  4. Set parameters: adjust any options the tool exposes; the paper itself describes a single-pass prediction with no sampling settings.

  5. Run the job: submit the run and wait for the prediction, which the paper reports takes a few seconds per complex.

  6. Download outputs: retrieve the docked complex, which is the ligand placed onto the fixed receptor through the predicted rotation and translation.

  7. Interpret and follow up: inspect interface contacts and clashes, and consider refining or re-ranking the pose with another method.

Things to Keep in Mind

  • EquiDock treats proteins as rigid, so it does not handle flexible docking such as many antibody cases.

  • The training datasets are biased toward rigid docking.

  • Steric clashes are discouraged only by a soft penalty, and there is no structure refinement or candidate re-ranking step.

  • On several metrics, other docking tools reported in the paper, notably HDock, were more accurate.

Source: Ganea, O.-E., Huang, X., Bunne, C., Bian, Y., Barzilay, R., Jaakkola, T., Krause, A. Independent SE(3)-Equivariant Models for End-to-End Rigid Protein Docking. arXiv:2111.07786. Code: https://github.com/octavian-ganea/equidock_public

Supporting 10,000+ scientists around the world,

from leading biotechs, and global biopharma