Use GeoDock Online
Commercially Available GeoDock No-Code Web Server
GeoDock: Fast, Flexible Protein-Protein Docking with a Multi-Track Iterative Transformer
GeoDock is a multi-track iterative transformer for fast protein-protein docking that predicts the structure of a complex from the unbound structures of its two partners. It uses ESM-2 (650M) sequence embeddings instead of multiple sequence alignments (MSAs), and includes invariant point attention so that the backbone is allowed to move. The method is described in a 2023 bioRxiv preprint.
Classical docking programs sample many poses and then score them, which is computationally heavy, and earlier deep learning docking methods were often limited to rigid-body prediction. GeoDock aims to give flexible, single-pass predictions that run in a fraction of a second.
Feature | Detail |
|---|---|
Task | Docking of binary protein complexes from unbound monomer structures |
Sequence input | ESM-2 (650M) embeddings, no MSA required |
Size | About 4.3M trainable parameters; 3-layer graph module and 3-layer structure module |
Training data | DIPS (36,318 train, 927 validation, 100 test) then fine-tuning on DB5.5 |
Success metric | DockQ of at least 0.23 (CAPRI acceptable) |
How GeoDock Predicts Docked Complexes
Build intra-chain graphs: Each chain is a fully connected graph with ESM-2 node embeddings and trRosetta-style distance and orientation edges. Inter-chain information is masked.
Iterate: The model runs 4 iterations, with recycling that includes inter-residue orientation distributions.
Move the backbone: Invariant point attention in the structure module lets the model update backbone frames.
Optional contact hints: During training, 0 to 3 random inter-chain contacts (C-beta distance under 10 Angstrom) were supplied; none were given at inference for blind docking.
Training combined frame-aligned point error, pairwise distance, lDDT and structure violation losses in three stages, pre-trained on DIPS and then fine-tuned on DB5.5.
GeoDock Accuracy and Speed Benchmarks
Benchmark | GeoDock result |
|---|---|
DIPS test set (blind docking) | 41% success, reported as the best of the tested methods |
DB5.5 bound set (24 complexes) | 25% success; above EquiDock and PatchDock, below ClusPro and ATTRACT |
DB5.5 unbound set (69 flexible targets) | 7% success versus 6% for ClusPro and 22% for ReplicaDock2 |
Inference time (one V100 GPU) | 0.76 s on DIPS, 0.80 s on DB5.5 |
The authors state GeoDock is at least a thousand times faster than the other methods compared. In their DsbC-DsbD example, backbone movement was under 1 Angstrom, so predictions stay close to the unbound inputs.
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 GeoDock on Tamarind Bio
Rapid complex modelling: Generate a docked pose for two proteins in about a second.
Large-scale screening: Triage many candidate protein pairs quickly, since inference is orders of magnitude faster than sampling-based docking.
Hint-guided docking: The authors designed the model around optional inter-chain contact information.
Starting poses: Use the output as an initial model for refinement or for comparison with slower docking methods.
How to Use GeoDock on Tamarind Bio
Log in: Sign in at tamarind.bio and open the GeoDock tool.
Provide the two partners: Supply the unbound structures or sequences of the two proteins as the tool requests.
Add contact hints (if exposed): Optionally provide known inter-chain contacts.
Run: Submit the job; the model runs a single forward pass with 4 iterations.
Download outputs: Retrieve the predicted complex structure.
Interpret: Compare the pose with known biology and check interface plausibility.
Use downstream: Refine with physics-based docking or use as a starting model for further analysis.
Things to Keep in Mind
Backbone flexibility is minimal because training examples contained no conformational change.
Success on unbound targets is low (7%), well below ReplicaDock2 (22%).
Success on DB5.5 (25%) is lower than on DIPS (41%), which is mostly homo-multimers.
It trails sampling-and-reranking methods because its training objective minimizes average error.
Only binary complexes were considered, and HDOCK and AlphaFold-Multimer were not compared because of leakage concerns.
Source: GeoDock, bioRxiv preprint, 2023, DOI 10.1101/2023.06.29.547134. No code repository was stated in the text consulted.