DiffDock-PP

When two proteins bind and neither deforms much, the entire answer is a rotation and a translation of one against the other — six numbers. A model trained to regress those six numbers under a mean-squared-error loss learns something subtly wrong: when several placements are plausible, the loss is minimised by predicting their weighted average, and an averaged pose is usually not a pose at all. This is the origin of the steric clashes in EquiDock's predictions, where the two chains are driven partly into one another. DiffDock-PP instead treats docking as having a distribution of answers, and builds a model that samples from it.

Rather than diffusing atom coordinates, DiffDock-PP puts a diffusion process on the product of the 3D translation group T(3) and the rotation group SO(3) — precisely the six degrees of freedom a rigid ligand has relative to a fixed receptor — and learns the score of that process. Inference randomly rotates and translates the smaller protein, runs reverse diffusion to draw candidate complexes, and hands them to a separately trained confidence model that predicts which are good. Because the diffusion lives on the pose manifold, every sample is a physically valid rigid placement by construction.

The model was built at the Technical University of Munich with collaborators at MIT CSAIL, and presented at the ICLR 2023 Machine Learning for Drug Discovery workshop. It carries the architecture of DiffDock, the MIT group's protein-small molecule docking model, over to protein pairs: torsional diffusion is dropped because the proteins stay rigid, and the two proteins share intra-protein convolution layers so the network respects the symmetry of a pair.

#Key Features

• Docking as sampling: Reverse diffusion produces distinct candidate complexes rather than a point estimate, sidestepping the averaging failure mode of regression-based docking and yielding clash-free structures.
• A learned pose ranker: The confidence model is a classifier trained on the score model's own samples to predict whether a pose falls below 5 Å ligand-RMSD, so predictions arrive with a score rather than as bare coordinates.
• Selective accuracy: Ordering predictions by confidence and discarding the least certain raises the success rate across the complexes that remain, making the model usable as a triage step over large candidate sets.
• Sequence context in the geometry: Each residue node carries an ESM-2 embedding alongside its residue type and alpha-carbon position, so evolutionary signal informs the geometric reasoning.

#Technical Details

Both the score model and the confidence model are SE(3)-equivariant tensor-field convolutional networks implemented with e3nn. Proteins are coarse-grained to one node per residue, connected to their 20 nearest intra-protein neighbours, while cross-protein edges use a cutoff of (40 + 3σ_tr) Å that shrinks as the translational noise decays. The score model emits two equivariant 3-vectors at the ligand centre of mass — the translational and rotational scores — while the confidence model mean-pools scalar features into a single invariant value. The whole system has 1.62M parameters and was trained for 170 epochs on the Database of Interacting Protein Structures (DIPS), 42,826 binary complexes under the protein-family-based split introduced with EquiDock. Inference uses low-temperature sampling to concentrate on high-likelihood modes.

On 100 held-out DIPS complexes, sampling 40 poses gives a median complex RMSD of 4.85 Å with 42% of predictions under 2 Å, and a median interface RMSD of 4.23 Å. HDOCK reaches a 6.23 Å median but needs 778 s per complex on CPU; AlphaFold-Multimer reaches 8.61 Å and EquiDock 13.30 Å. A single sample takes 4.2 s and lands at 11.95 Å; the 40-sample run takes 153 s. Oracle selection over those 40 gives a 0.67 Å median with 71% under 2 Å, locating most of the remaining error in ranking rather than in the generative model.

#Applications

Rigid docking is the workhorse step for anyone holding structures for two partners and needing the complex: mapping interaction interfaces, prioritising candidate protein-protein interactions for experimental follow-up, and antibody-antigen modelling where the two chains are known separately. The GPU runtimes make it practical at screening scale, where thousands of pairs must be triaged and the confidence score decides which deserve slower treatment. The repository also writes the reverse diffusion trajectory out as PDB files for inspection in PyMOL.

#Impact

DiffDock-PP established that the generative framing that worked for small-molecule docking transfers to protein-protein complexes, and it is now a standard baseline in the protein-docking literature — later diffusion and flow-matching docking models routinely report against its DIPS numbers. Its limitations are equally clear. It is a workshop paper rather than a peer-reviewed article; the released checkpoints were trained on DIPS only, with DB5.5 left unevaluated in the paper; and the rigid assumption models no conformational change on binding. The checkpoints ship as loose files in the git tree with no model card, and while the README states MIT, the repository carries no license file.