Loading…
Blind protein-ligand docking that transfers to binding domains absent from training, scoring 22.6% top-1 on DockGen and 50% on PoseBusters.
No providers recorded yet. Browse all providers
The PDB entries 1QXZ and 5M4Q share 22% sequence identity — below the 30% cutoff a similarity split uses to call two proteins unrelated — yet their binding pockets superimpose almost exactly. Pockets are among the most conserved parts of a protein, so splitting docking data by sequence leaves the test set full of binding modes the model has memorized. Classify the same complexes by the protein domain the ligand binds, using the ECOD hierarchy, and the roughly 17,000 pre-2019 PDBBind structures collapse into 487 clusters, with the 2019 test set adding only eight more. Measured that way, blind docking accuracy looks very different: DiffDock reaches 38.2% top-1 under 2 Å on the PDBBind time split and 6.0% on ligand-binding domains it has never seen.
DiffDock-L is the model built to close that gap, from MIT CSAIL with collaborators at UC Berkeley and the Dana-Farber Cancer Institute and Harvard Medical School, presented at ICLR 2024. It keeps the DiffDock formulation — diffusion over ligand translation, rotation and torsion angles, with a confidence model ranking the sampled poses — and changes the three things a scaling study said mattered: a larger score model, a training set extended with Binding MOAD, and a synthetic task that manufactures binding pockets from protein sidechains.
The same paper introduces Confidence Bootstrapping, a self-training scheme in which a diffusion model's rollouts are scored by the confidence model and fed back into the early denoising steps. That is a training procedure applied to a small variant, DiffDock-S, not a step users of the released checkpoint run.
Both networks are SE(3)-equivariant tensor-field convolutional networks over heterogeneous point clouds, with receptor residues carrying cached ESM-2 650M embeddings. Training combined the PDBBind pre-2019 split, the Binding MOAD complexes from domains already represented there, and van der Mer synthetic complexes drawn from the ProteinMPNN structure set. On PDBBind, ten samples give 43.0% top-1 under 2 Å at a median RMSD of 2.8 Å in about 25 seconds per complex, against 38.2% for DiffDock with 40 samples. On DockGen-full the figure is 22.6% at a median 4.3 Å, and 27.6% on the DockGen-clusters subset. On the PoseBusters set, 50% of top-1 poses fall under 2 Å, against 38% for DiffDock and 42% for RoseTTAFold-All-Atom, while pocket-conditioned Vina reaches 60% with the site handed to it. The score and confidence checkpoints ship as a GitHub release asset under an MIT license and download automatically on first run.
The gain is concentrated where blind docking is worth doing: targets whose pockets are not represented in crystallographic databases. That covers reverse screening a compound against many proteins to anticipate off-target effects, proposing substrate poses for enzymes and membrane proteins of unknown function, and generating starting geometries for physics-based rescoring. The confidence score reports pose quality, not binding affinity, so affinity questions still call for a scoring function or a free-energy calculation on the relaxed pose.
DockGen has become a standard test of whether a docking model learned binding physics or memorized pocket families, and DiffDock-L is what the repository runs by default, with the original reachable at the v1.0 tag. Confidence Bootstrapping is the more open-ended contribution: fine-tuning DiffDock-S on individual unseen domains without structural data for them raised DockGen-clusters accuracy from 9.8% to 24.0%, above 30% in half the domains. Limitations carry over from the original: evaluation uses holo receptor structures, so the protein is assumed already in its bound conformation, ligands are drug-like small molecules, and protein-protein and protein-nucleic-acid docking are out of scope. The name also travels — NVIDIA serves a DiffDock microservice built on this architecture but trained on its own complexes under a separate license, and its numbers are not these.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.