Johns Hopkins University / GrayLab
A generative protein-protein docking model that adapts AlphaFold-Multimer via flow matching, replacing the template module with a docking module.
AF2Dock is a generative model for structure-based protein–protein docking that adapts a co-folding model—AlphaFold-Multimer (AF-M)—into a docking engine. Instead of co-folding two chains from sequence and multiple sequence alignments (MSAs), AF2Dock takes the individual subunit structures as input and predicts how they assemble. It does this by replacing AF-Multimer's template module with a new docking module and retraining the model end-to-end with a flow-matching objective, so the network learns to transport an initial pose distribution toward correct complex geometries.
Developed in Jeffrey Gray's lab at Johns Hopkins University and posted to bioRxiv in late 2025 (with a revised v3 in April 2026), AF2Dock sits at the intersection of two trends: leveraging powerful pretrained co-folding backbones, and using flow-matching/diffusion-style generative training for sampling docked poses. Because it docks from given structures rather than folding from sequence, it does not require MSAs at inference, distinguishing it from standard AF-Multimer and AlphaFold3 usage.
A central finding is that AF2Dock produces predictions that are orthogonal to co-folding models: it succeeds on some targets where AF-Multimer and AF3 fail, making it a complementary tool rather than a strict replacement. It is most notable for antibody-related complexes—on nanobody complexes it outperforms all other docking methods the authors tested.
AF2Dock is built on the OpenFold reimplementation of AlphaFold-Multimer. The template module is replaced by a docking module, and the AF-M components are trained with a flow-matching objective on the PINDER protein–protein interaction dataset, with support for holo, apo, and predicted structure combinations as inputs. It was evaluated on the PINDER-AF2 benchmark and a dedicated antibody/nanobody test set. Using non-holo (apo or predicted) inputs, AF2Dock is competitive with baseline docking methods and outperforms all tested methods on nanobody complexes; on general complexes it trails co-folding AF-Multimer and AF3 in raw success rate but yields orthogonal correct predictions. Ablations show that full-parameter fine-tuning of the AF-M components is critical, and—counterintuitively—that adding ESM embeddings can reduce success on some targets, including nanobodies. Pretrained weights are distributed via Zenodo (10.5281/zenodo.17782958).
AF2Dock is aimed at researchers predicting protein complex structures when subunit structures are already available, such as docking antibodies or nanobodies onto antigens for therapeutic and immunology research, assembling experimentally or computationally determined components, and modeling interactions where MSAs are unavailable or uninformative. Its orthogonality to AF-Multimer and AF3 makes it especially useful inside ensemble pipelines, where sampling from multiple methodologically distinct predictors increases the chance of recovering a correct pose for difficult targets.
AF2Dock demonstrates that large pretrained co-folding models can be repurposed into generative docking engines through flow-matching fine-tuning, offering an MSA-free alternative that is complementary to AlphaFold-Multimer and AlphaFold3. Its strong performance on nanobody complexes is significant given the importance of single-domain antibodies in therapeutics and the difficulty co-folding methods often face on them. With open code (MIT) and weights on Zenodo, the work is readily reproducible and extensible. As a preprint, its broader benchmarks are still maturing, and the model inherits both the strengths and the computational footprint of the AlphaFold-Multimer backbone it builds on.