Loading…
Rigid protein-protein docking model that predicts a complex from two unbound structures in one pass, with no candidate sampling or refinement.
Rigid protein-protein docking asks a narrow question: given two proteins whose individual structures are already known, where does one sit relative to the other once they bind? Established docking software answers it by brute force — sample millions of candidate placements, rank them with a scoring function, then refine the survivors with an energy model or force field. That takes minutes to hours for a single pair and still carries no guarantee of landing on the right interface. EquiDock replaces the whole pipeline with one forward pass that predicts the rotation and translation placing one protein (the ligand) against the other (the receptor), using no candidate sampling, no re-ranking, no templates and no structure refinement.
What makes a direct prediction tractable is a symmetry constraint imposed by construction. A docked complex should not depend on where the two input structures happen to sit in their coordinate files, nor on which of the two an operator labelled "ligand". EquiDock builds in both: the network is pairwise-independent SE(3)-equivariant and commutative, so the predicted complex is identical up to a rigid transformation no matter how the inputs are positioned, rotated, or which role each is assigned. Prior equivariant networks handled a single point cloud; the paper derives the necessary and sufficient constraints for a pair of independently placed 3D objects and designs the architecture to satisfy them, which removes the need for rotational data augmentation entirely.
EquiDock was developed by Octavian-Eugen Ganea and colleagues at MIT CSAIL, ETH Zurich and Tencent AI Lab, and published at ICLR 2022 — the same MIT group that went on to build DiffDock for protein-ligand docking.
The architecture is an Independent E(3)-Equivariant Graph Matching Network (IEGMN) operating on k-nearest-neighbour residue graphs built from alpha-carbon coordinates, interleaving graph feature embedding updates with equivariant coordinate updates. Released configurations use five to eight IEGMN layers at hidden dimension 64, so the distributed checkpoints are only a few megabytes; the papers report no parameter count. Training uses the Database of Interacting Protein Structures (DIPS), filtered to proteins of at most 10K atoms and split 39,937/974/965, with Adam at learning rate 2e-4 and early stopping after 30 epochs of patience. A second checkpoint fine-tunes that model on Docking Benchmark 5.5 (253 structures, split 203/25/25) at learning rate 1e-4. On the DIPS test set EquiDock reaches a median complex RMSD of 13.29 Å and median interface RMSD of 10.18 Å; on DB5.5 the medians are 14.13 Å and 11.97 Å. Those figures are competitive with, and on several measures better than, ATTRACT, HDOCK, ClusPro and PatchDock — with the caveat the authors raise themselves, that the proprietary baselines may have seen parts of the test sets during their own validation.
The combination of fixed pretrained weights and single-pass inference suits work where docking must run many times rather than once: screening candidate interaction partners across a proteome, generating structural hypotheses for protein pairs with no experimental complex, or serving as a fast differentiable docking module inside a larger design loop such as binder or antibody generation. The rigid-body assumption is the boundary — proteins that reorganize substantially on binding, antibodies among them, are outside what the model represents.
EquiDock established pairwise-independent equivariance as a design principle for docking networks and became a standard reference point for the learned dockers that followed, including DFMDock and DockFormer. Both author checkpoints remain distributed through NVIDIA's BioNeMo collection on NGC as NeMo-format conversions of the published weights rather than retrained models, which is how most current users encounter it. The original repository is unmaintained following the death of its lead author in 2022, and pins a 2022 dependency stack with hard-coded input paths, so running it on new structures takes some editing.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.