DFMDock

DFMDock is a diffusion-based model for protein-protein docking developed by Lee-Shin Chu, Sudeep Sarma, and Jeffrey Gray at Johns Hopkins University. Predicting how two proteins assemble into a complex requires both generating plausible bound poses and reliably identifying which pose is correct. Most recent diffusion-based docking methods decouple these two problems, sampling structures with one model and ranking candidate decoys with a separate scoring function. DFMDock unifies them in a single framework.

The model jointly learns to sample structures and to rank them by predicting denoising forces alongside a scalar energy value, supervised through physically motivated objectives that combine force matching with contrastive training. Because the energy is learned as part of the same network that generates poses, DFMDock produces an internally consistent ranking signal rather than relying on an external scoring tool. A notable practical property is that it operates without multiple sequence alignments and generalizes to targets not seen during training.

Introduced on bioRxiv in September 2024 and updated in 2025, DFMDock is released with training and inference code and pretrained weights, following a line of generative approaches to macromolecular docking.

#Key Features

• Unified sampling and ranking: A single model predicts both denoising forces for pose generation and a scalar energy for ranking, avoiding the separate scoring pipelines used by prior diffusion dockers.
• Physically motivated supervision: Training combines a force-matching objective with contrastive learning, encouraging the learned energy surface to distinguish near-native poses from decoys.
• MSA-free operation: Docking is performed without multiple sequence alignments, simplifying inference and broadening applicability to targets with shallow alignments.
• Generalization to unseen targets: The model transfers to complexes outside its training set, an important property for prospective docking on novel protein pairs.
• Learned energy function: The energy predicted by DFMDock outperforms Rosetta energy and model-derived confidence scores as a ranking criterion.

#Technical Details

DFMDock casts rigid-body docking as a denoising diffusion process over the relative pose of two protein partners. The network is trained to predict the forces that reverse the diffusion corruption and, simultaneously, a scalar energy assigned to each configuration; force matching ties the generative dynamics to the energy landscape, while contrastive training sharpens the energy's ability to rank candidate poses. On Docking Benchmark 5 (DB5), DFMDock reaches a 32.8% Oracle success rate and a 5.3% Top-1 success rate, improving on DiffDock-PP, which reports 16.2% and 4.3% respectively under the same evaluation. The authors further show that the learned energy is a stronger ranking signal than Rosetta energy or confidence scores derived from structure predictors. Code, pretrained weights, and DIPS-based checkpoints are distributed through the Gray Lab GitHub repository under the MIT license; the repository does not include a formal model card.

#Applications

DFMDock is aimed at structural biologists and computational researchers who need to predict the three-dimensional structure of protein-protein complexes, particularly when deep multiple sequence alignments are unavailable or when candidate poses must be ranked without external energy functions. Because it both samples and scores, it can be used to enumerate plausible binding modes and prioritize them in a single pass, supporting interactome modeling, mechanistic studies of complex formation, and hypothesis generation for interactions lacking experimental structures.

#Impact

DFMDock illustrates that generative sampling and energy-based ranking need not be separate stages: coupling them through force matching and contrastive supervision yields a learned energy that ranks poses more effectively than established physics-based and confidence-based alternatives. The roughly two-fold improvement in Oracle success rate over DiffDock-PP on DB5 marks meaningful progress for MSA-free diffusion docking. As a preprint benchmarked in silico, its performance on prospective, experimentally validated targets remains to be established, and success rates on DB5 indicate that rigid-body docking of arbitrary complexes remains an open challenge.