Use AF2Dock Online
Commercially Available AF2Dock No-Code Web Server
AF2Dock: Flow-Matching Protein-Protein Docking Built on AlphaFold-Multimer
AF2Dock is a generative protein-protein docking model that adapts the AlphaFold-Multimer (AF-M) co-folding network and trains it end-to-end with flow matching. It was developed by Da Xu, Lee-Shin Chu and Jeffrey J. Gray in the Department of Chemical and Biomolecular Engineering at Johns Hopkins University, and is described in the bioRxiv preprint "Adapting Co-Folding Models for Structure-Based Protein-Protein Docking Through Flow Matching". Code is available on GitHub, and model weights are deposited on Zenodo.
Co-folding models such as AlphaFold rely on multiple sequence alignments (MSAs), which limits them on hard targets such as antibody-antigen complexes. AF2Dock instead starts from the unbound partner structures and learns to dock them, replacing the AF-M template module with a docking module that denoises a noisy complex structure.
Feature | Detail |
|---|---|
Base model | AlphaFold-Multimer (OpenFold implementation) with a docking module in place of the template module |
Generative method | Conditional flow matching; 10 integration steps by default |
Training data | PINDER training set, about 2.3 million bi-protein complexes |
Input monomers | Holo, apo or predicted structures, plus per-chain ESM-C embeddings |
Ranking | Predictions ranked by ipTM |
How AF2Dock Works
Docking module: It embeds the noisy complex structure using a distogram, backbone-frame unit vectors and masks, and a diffusion-transformer-style pair denoiser with adaptive layer normalization conditioned on time. The result feeds the Evoformer.
Flow matching: The network predicts the fully denoised structure along a conditional optimal-transport path, starting from a rigid-body prior with random translation and rotation.
Flexible inputs: During training, each chain is sampled as holo, apo or predicted structure with probabilities of 20%, 40% and 40%.
Three training stages: The docking module is trained alone on holo monomers, then fine-tuned with AF-M frozen, then all parameters are fine-tuned together.
Sampling and ranking: Multiple structures are sampled per target (20 for PINDER-AF2, 40 for antibody and nanobody sets) and ranked by ipTM.
Recycling is disabled because the iterative flow-matching steps play a similar role.
Benchmark Results for Antibody, Nanobody and General Docking
Benchmark | Test set | Reported outcome |
|---|---|---|
PINDER-AF2 | 180 holo, 30 apo, 127 predicted monomer inputs | Competitive with or better than other docking methods on apo and predicted inputs; below co-folding AF-M and AF3 |
Nanobody-antigen | 60 complexes, predicted inputs | Outperforms all tested docking methods, including HDock, ZDock and single-sequence AF-M |
Antibody-antigen | 49 complexes | Favours the base AF2Dock over the non-ESM variant |
Success is measured with DockQ thresholds (acceptable above 0.23, medium above 0.49, high above 0.80) at top-1, top-5 and oracle levels. AF2Dock underperforms co-folding AF-M and AF3 on success rate with non-holo inputs, but its correct predictions are largely orthogonal to theirs, so it can find correct structures where co-folding fails. Ablations show full-parameter fine-tuning of AF-M is critical, and ESM embeddings can hurt in some cases, notably nanobody complexes. The paper text does not give the percentages, which appear in its figures.
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 AF2Dock on Tamarind Bio
Nanobody-antigen docking: Model nanobody complexes from separately predicted or known structures.
Antibody-antigen docking: Generate candidate poses for antibody-antigen pairs, where MSA-dependent methods struggle.
Complementary pose sampling: Compare AF2Dock poses with co-folding predictions, since the authors report orthogonal successes.
Docking from unbound or predicted structures: Dock apo or predicted monomers rather than requiring holo structures.
How to Use AF2Dock on Tamarind Bio
Log in: Sign in at tamarind.bio and open the AF2Dock tool.
Provide the partners: Supply the two protein chains as sequences and structures, as the tool requests; the authors docked two bodies.
Choose the variant: Select the standard or non-ESM variant if exposed; the paper found the non-ESM variant better on nanobodies but the base model better on antibodies.
Set sampling options: Adjust the number of samples or integration steps if exposed; the study used 10 steps and 20 to 40 samples.
Run: Submit the job.
Download and rank: Retrieve the sampled structures and rank them by ipTM.
Interpret: Check several top poses, since ipTM can be overconfident on false positives.
Things to Keep in Mind
AF2Dock does not explore internal flexibility effectively, and its performance drops most on difficult targets.
ipTM is overconfident for false positives, so naive re-ranking with co-folding models did not reliably help.
The antibody and nanobody benchmark used less stringent filtering, so data leakage is possible.
Only two-body docking was evaluated, and the apo set has just 30 targets.
Predicted CDRH3 loops often differ from the bound structure, causing clashes.
The license is not stated in the text consulted; check the repository before reuse.
Source: Xu D, Chu L-S, Gray JJ. Adapting Co-Folding Models for Structure-Based Protein-Protein Docking Through Flow Matching. bioRxiv preprint (2025/2026), DOI 10.1101/2025.11.28.691195. Code: github.com/Graylab/AF2Dock. Weights: Zenodo 17782958.