DiffDock

Ask a regression-trained docking model to place a ligand in a protein with two equivalent pockets and it puts the molecule between them: squared error is minimized at the mean of the plausible answers, not at any of them. The same averaging collapses atoms onto each other inside the ligand — 26% of top-1 poses from the regression model EquiBind clash sterically with the receptor. Docking is a one-to-many problem, and a model that must emit one point estimate has no way to say so.

DiffDock, from MIT CSAIL and published at ICLR 2023, reframes docking as generative modeling: learn a distribution over ligand poses conditioned on the protein, then sample it. What makes that tractable is a change of coordinates. A pose looks like 3n atomic positions, but bond lengths and angles are stiff enough to hold fixed, so any pose reachable from a seed conformer is pinned down by the ligand's position, its orientation, and its m torsion angles. DiffDock diffuses over that product space — translations, rotations, and one circle per rotatable bond, m + 6 dimensions instead of 3n — where the kernel can be sampled directly and every sample is chemically valid by construction.

Sampling alone does not yield one answer, so a second network — the confidence model — scores each pose for whether it lands within 2 Å of the truth and ranks them. The pair sits between brute-force search and one-shot prediction: many hypotheses compared, no high-dimensional search.

#Key Features

• No binding pocket required: The translational part of the diffusion finds the site, so the inputs are just a protein structure — or a bare sequence, folded on the fly with ESMFold — plus a SMILES string or an SDF file.
• Diffusion over degrees of freedom: Denoising acts on translation, rotation and torsion angles rather than Cartesian coordinates, keeping bond geometry intact: 2.8% of top-1 poses show a steric clash, against 26% for EquiBind and 6.6% for TANKBind.
• Confidence model that ranks and filters: Trained on whether a sampled pose falls below 2 Å RMSD, it tracks pose quality at Spearman 0.68; on the most confident third of complexes, success rises from 38% to 83%.
• Language-model protein features: Receptor residues carry cached ESM-2 650M embeddings. The score model sees a coarse-grained α-carbon graph, the confidence model an all-atom one.

#Technical Details

Both networks are SE(3)-equivariant tensor-field convolutional networks over heterogeneous point-cloud graphs. The score model has 20.24 million parameters over six convolution layers; the confidence model has 4.77 million over five. Training used the PDBBind time split: about 17,000 complexes from 2018 or earlier, 363 test structures from 2019, and 850 epochs on four 48 GB A6000 GPUs. On PDBBind blind docking, top-1 success at RMSD below 2 Å is 38.2% with 40 samples and 35.0% with 10, and top-5 is 44.7%, against 22.9% for GNINA, 21.8% for GLIDE, 20.4% for TANKBind and 5.5% for EquiBind. Docking into ESMFold-predicted structures rather than crystal ones gives 21.7% top-1 where the best baseline manages 10.4%. Ten samples take about 10 seconds on an A100, three to twelve times faster than GNINA.

Three artifacts now carry the DiffDock name, and their licenses, training sets and benchmark numbers do not transfer between them. This entry describes the original 2022 model, whose score and confidence checkpoints ship in gcorso/DiffDock under an MIT license; since February 2024 that repository runs the successor DiffDock-L by default, and the original is reached with git checkout v1.0. NVIDIA separately serves a v2.2 NIM microservice that reuses the MIT architecture but whose weights NVIDIA trained on 486,000 complexes drawn from PLINDER and SAIR, released under the NVIDIA Open Model License rather than MIT.

#Applications

The intended use is structure-based drug discovery where the binding site is unknown or uncertain: screening compound libraries against a target, reverse screening one compound against many targets, and generating starting poses for physics-based scoring. Because the protein can be supplied as a sequence, DiffDock fits pipelines built on predicted rather than experimental structures. Its confidence score measures pose quality, not binding affinity; the authors recommend pairing it with a scoring function or a free-energy calculation when affinity is the question.

#Impact

DiffDock established diffusion over a pose manifold as the standard formulation for learned docking, and it is the reference baseline later docking models measure themselves against — PhysDock, DTMol, ArtiDock, CWFBind and DFMDock among them. The limitations are worth stating plainly: training and evaluation used holo crystal structures, so the model assumes a receptor already in its bound conformation and treats flexibility only implicitly; the PDBBind time split has since been criticized for leaving similarity between train and test, motivating the harder DockGen benchmark; and it is built for drug-like small molecules, with protein-protein and protein-nucleic-acid docking directed elsewhere.