DeltaDock

Blind docking asks a method to find the binding pocket and place the ligand inside it at once. The pocket half is already well served: Fpocket, P2Rank, SiteMap, DoGSite3 and DSDP each recover the true site 70–80% of the time, and pooling their predictions pushes recall to nearly 95%. The bottleneck is not a better pocket predictor but a way to pick the right pocket, for a particular ligand, out of that pool.

DeltaDock, from the State Key Laboratory of Cognitive Intelligence at the University of Science and Technology of China and the Center for Quantitative Biology at Peking University, builds on that observation. It reframes pocket prediction as a pocket-ligand alignment problem: a contrastive module embeds the query ligand and every candidate pocket into a shared space and picks the pocket closest to the ligand. A second stage then places the ligand with a bi-level coarse-to-fine refinement network. Splitting the work this way lets one framework serve both regimes — supply a pocket and the first stage is skipped.

The name describes that second stage: poses come from iteratively refining an input structure rather than being generated from scratch. Geometric deep learning docking moves atoms directly and can return distorted bond geometry or steric clashes — DiffDock passes only 14% of the PoseBusters validity checks against a 38% raw success rate — so DeltaDock, presented at NeurIPS 2024, adds steric and mobility penalties to training and a fast post-processing step that restores plausible geometry.

#Key Features

• Pocket selection as contrastive alignment: CPLA encodes the ligand with Attentive-FP and each candidate pocket with a geometric vector perceptron over the residue graph, then ranks pockets by cosine similarity to the ligand embedding.
• Bi-level structure refinement: Bi-EGMN refines residues inside a 40 Å cubic box around the pocket, then switches to atom-level detail within 6 Å of the ligand — side-chain resolution for close contacts without full-atom cost across a large pocket.
• Physics-informed objective: Van der Waals and bound-matrix constraint terms sit beside the distance-map losses, penalising clashes and implausible ligand geometry during training.
• Fast structure correction: Predicted coordinates are re-expressed as torsion angles of the input conformer with RDKit and, on a detected clash, minimised with SMINA — about 0.4 seconds per pair, against minutes for OpenMM.

#Technical Details

Candidate pockets come from DSDP and P2Rank; adding Fpocket degraded contrastive training by supplying weak negatives. CPLA's residue features are drawn from the ESM-2 3B protein language model, and it trains in about two hours on a single A100-40GB. Bi-EGMN is an E(3)-equivariant graph matching network initialised from the GPU-accelerated pose sampler introduced with DSDP and recycled four times at inference; it takes roughly 48 hours on four A100s. Both stages were trained on PDBbind v2020 under a time split — 17,299 training, 968 validation and 363 test complexes, with everything from 2019 onward held out.

Blind docking on the PDBbind test set places 47.4% of ligands within 2 Å RMSD against 36.0% for DiffDock, the 31% relative improvement the paper headlines, at about 2.97 seconds per pair. On the 142-complex unseen subset it reaches 40.8% against 19.4% for the strongest deep-learning baseline. On the 428-complex PoseBusters benchmark it reaches 48.8% when physical validity is required jointly with the RMSD threshold; that comparison, not the RMSD-only one, is where the roughly 300% figure comes from. Ablations credit both stages: removing CPLA drops PDBbind accuracy to 41.2% and removing Bi-EGMN to 44.6%.

#Applications

The framework fits structure-based drug design on targets whose binding site is uncertain or whose pocket is too large for the 12 Å truncation typical of site-specific methods — the multi-sub-pocket SARS-CoV-2 main protease is the paper's worked example. Its per-pair cost suits triaging a screening library, and the same checkpoint serves a campaign that later acquires a co-crystal structure and switches to site-specific docking. Poses arrive filtered for steric and geometric plausibility, cutting the cleanup normally needed before free-energy calculation or medicinal chemistry review.

#Impact

The durable contribution is the reframing: given a mature ensemble of pocket finders, selection conditioned on the ligand is a more tractable target than another pocket predictor, and treating physical validity as a training objective rather than a post-hoc filter narrows the gap between deep-learning docking and the physics-based programs it aims to replace. Two constraints bound its reach: the receptor is held rigid, so induced-fit effects are out of scope, and the pipeline leans on DSDP, P2Rank, RDKit and SMINA rather than standing alone. The larger obstacle is access. The paper states that code and data will be released at the project's GitHub repository, but that repository holds only a README, a conda environment script and a figure, and has not been updated since October 2024. Running DeltaDock means reimplementing it.