Protein-ligand docking model for high-throughput virtual screening, predicting binding poses with graph neural networks at low computational cost.
No providers recorded yet. Browse all providers
Virtual screening imposes an unusual constraint on a docking method: it must place millions of candidate ligands into a known pocket, so per-ligand cost matters as much as per-ligand accuracy. Classical physics-based programs such as AutoDock, Vina, and Glide are cheap enough to run at that scale but have plateaued in accuracy, while the machine-learning docking wave that followed largely optimised for blind docking or co-folding — regimes that are slower and answer a question screening campaigns do not ask, since the pocket is usually already known.
ArtiDock, from Receptor.AI, targets the screening regime directly. It performs pocket-specific rigid docking with a graph neural network that predicts the pocket-ligand distance matrix, from which a 3D pose is reconstructed and aligned to a ligand conformer. There is no diffusion sampling and no structure co-folding, which is what keeps inference cheap: the authors estimate roughly $5 of compute to dock a million molecules, against $12 for AutoDock-GPU and $15 for Glide.
First described in a March 2024 preprint and substantially revised in May 2025, the work appeared in the Journal of Chemical Information and Modeling in January 2026. Alongside the model, the authors argue for a benchmark redesign, building their evaluation on the PLINDER dataset to mimic the industrial screening scenario rather than the curated, redocking-friendly test sets common in the literature. ArtiDock 2.5 is the version deployed inside Receptor.AI's commercial drug design platform; external access is by request rather than through a public checkpoint.
Ligands are encoded as atom-level molecular graphs through a graph isomorphism operator, while the protein pocket's heavy atoms are encoded with a Geometric Vector Perceptron GNN for SE(3) equivariance. Training used the PLINDER 2024-06 v2 split, filtered to 227,066 protein-ligand pairs drawn from 206,555 systems and 67,888 PDB entries with 29,440 distinct chemical components; these were grouped into 40,148 clusters by ligand and pocket similarity, with one pair sampled per cluster per epoch to limit the influence of overrepresented ligands.
On the PLINDER holo test set of 675 systems, ArtiDock reaches 48% of poses under 2 Å RMSD with a median of 2.0 Å and lDDT-PLI of 0.71, against 39% for Glide, 38% for Vina, and 35% for AutoDock-GPU — the 29–38% relative accuracy gain the paper headlines. Chemical validity is the trade-off: 72% PoseBusters-valid out of the box versus 91–95% for the physics-based methods, rising to 95% after UFF minimisation at the cost of dropping to 41% under 2 Å. On the time-split PoseX benchmark, ArtiDock reaches 73.5% under 2 Å for self-docking (SurfDock: 77.6%) and 62.4% for cross-docking when PoseBusters validity is jointly required, ahead of the cofolding model Boltz-1x, a physically refined variant of Boltz-1, at 60.2%.
The intended user is a computational chemistry team running large virtual screens against a structurally characterised target, where the pocket is known and throughput dominates. The method's handling of bridging waters and metal ions makes it a candidate for target classes where those features mediate binding, such as metalloenzymes and kinases with conserved water networks. Because poses are not scored intrinsically, it slots in as the geometry stage of a pipeline that supplies its own affinity or rescoring model downstream.
ArtiDock is one of the clearer demonstrations that machine-learning docking can be both more accurate and cheaper than classical programs when scoped to the pocket-specific screening task rather than to blind docking. Its accompanying PLINDER-based benchmark, released as a public evaluation harness, is a contribution in its own right, since it exposes how much reported docking accuracy depends on test-set construction. The main constraint on adoption is openness: the model weights are proprietary, no public checkpoint or open API exists, and the released repository contains the benchmark and classical-docking setup rather than ArtiDock inference. Independent replication therefore rests on the published numbers, and the reported chemical validity gap means poses need physics-based cleanup before downstream use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.