Use AF2Dock Online
Commercially Available AF2Dock No-Code Web Server
AF2Dock: Structure-Based Protein-Protein Docking Powered by Flow Matching
Predict bound protein complex structures from unbound monomers without multiple sequence alignments (MSAs).
What is AF2Dock?
AF2Dock is a novel generative docking model designed to adapt advanced co-folding architectures for structure-based protein-protein docking. Traditional co-folding frameworks like AlphaFold-Multimer (AF-M) rely heavily on Multiple Sequence Alignments (MSAs) to infer inter-residue contacts, which often severely limits their accuracy on challenging evolutionary targets like antibody-antigen or pMHC-TCR complexes.
By replacing the template module of OpenFold (an AF-M implementation) with a dedicated docking module and training end-to-end with a flow-matching objective, AF2Dock operates successfully in single-sequence mode without needing MSAs. It works directly on noisy or unbound monomer structures to iteratively denoise them into the accurate bound complex structure.
Key Performance Benefits
State-of-the-Art Generative Performance: Substantially outperforms existing diffusion-based docking models like DiffDock-PP and DFMDock, dramatically bridging the performance gap with traditional physics-based tools.
Orthogonal Predictions to Co-Folding Models: Even when co-folding architectures like AF-M and AlphaFold 3 (AF3) fail on certain targets due to weak or deceptive sequence-based signals, AF2Dock produces entirely orthogonal predictions, successfully identifying correct structures where traditional models stumble.
Unrivaled Nanobody Performance: When evaluated using non-holo (unbound/predicted) inputs, AF2Dock outperforms all other tested structure-based docking methods on nanobody complexes.
Inpainting Capabilities: Inheriting the core strengths of the AlphaFold architecture, AF2Dock can selectively "inpaint" missing residues or highly flexible regions (like CDRH3 loops in antibodies), enhancing conformational diversity and structural alignment.
How It Works
The Docking Module: The core innovation replaces AlphaFold’s template module. Instead of simply embedding static data, it processes a noisy structure input along with positional features, transforming them into a denoised pair representation via a specialized diffusion-transformer pair denoiser stack.
Conditional Flow Matching (CFM): Trained with a CFM objective, AF2Dock learns to compute a velocity field that transforms a rigid-body prior and internal flexibility prior distribution into the precise bound data distribution over 10 iterative inference time steps.
Refined Ranking: Multiple structural conformations are generated by starting from different initial noises. Each prediction is accurately scored and ranked using the model's confidence metrics, such as the ipTM score.
What is Tamarind Bio?
Tamarind Bio is a premier computational biology platform engineered to democratize access to state-of-the-art deep learning algorithms for structural biology, bioinformatics, and molecular design. We remove the barriers of complex command-line setups, massive dependency installations, and expensive hardware procurement. By hosting advanced architectures on powerful, managed cloud GPUs, Tamarind Bio enables researchers, biochemists, and developers to run cutting-edge models seamlessly via an intuitive web interface or developer-friendly APIs.
How to Use AF2Dock on Tamarind Bio
Running structure-based docking with AF2Dock on Tamarind Bio is simple and requires no local installation or command-line scripting:
Upload Your Monomers: Provide individual protein sequences and their corresponding unbound structural configurations (such as apo structures, holo structures, or single-chain alphafold-predicted structures) in standard PDB format.
Configure Inference Settings: * Set your number of generation steps (default is 10 time steps).
Choose your sampling quantity (e.g., sample 20 structures for general benchmarks or up to 40 structures for complex antibody-antigen targets).
(Optional) Opt to run the Non-ESM Variant if you are docking smaller interface structures like nanobody-antigen complexes.
Submit the Run: Tamarind Bio will allocate high-performance cloud GPUs to process the end-to-end flow-matching pipeline.
Analyze and Download: Inspect the interactive 3D visualizations directly on your dashboard. Filter and rank your structural ensembles by their internal ipTM confidence scores, and download the refined PDB coordinate files for downstream analysis.
Frequently Asked Questions
Can AF2Dock model full protein flexibility?
AF2Dock explicitly accounts for internal protein flexibility during its conditional flow-matching training. However, like most structure-based deep learning docking tools, it exhibits minimal internal structural movements if the target requires massive binding-interface conformational changes. For targets with significant loop movements, combining AF2Dock with localized physics-based relaxation tools or utilizing the missing loop inpainting strategy on Tamarind Bio is highly recommended.
When should I use the Non-ESM variant of AF2Dock?
Ablation studies reveal that while ESM sequence embeddings can help identify larger antibody epitopes, they can introduce spurious signals that hinder success rates on smaller interfaces. If you are working specifically with nanobody complexes, we recommend selecting the non-ESM variant on Tamarind Bio, as it significantly improves Top-1 and Top-5 docking success rates in this setting.