Pocket-based protein-ligand pose predictor fine-tuned from pretrained Uni-Mol encoders, built to produce poses that pass physical plausibility checks.
No providers recorded yet. Browse all providers
A docked pose can be close to the crystal structure by RMSD and still be chemically impossible: an inverted stereocentre, a ligand atom inside a protein side chain, a bond stretched past anything real. The PoseBusters study showed that deep-learning docking methods often won on the RMSD metric while failing exactly these checks, and that their headline accuracy fell on complexes released after their training data. Uni-Mol Docking V2 is an answer to both criticisms from the group behind Uni-Mol.
Released in May 2024 by DP Technology and the AI for Science Institute in Beijing, it keeps the docking design of the original Uni-Mol paper and changes what surrounds it. The model starts from the two self-supervised Uni-Mol checkpoints, one pretrained on molecular conformers and one on protein pockets, and is fine-tuned end to end on protein-ligand complexes. The authors first rebuilt the data preparation for the original docking head, standardizing how pockets are cut and inputs generated, which alone lifted its PoseBusters accuracy from the 22% reported by the benchmark's authors to above 62%. V2 then retrains on a larger, carefully prepared corpus.
The task is pocket-conditioned re-docking: the user supplies a protein structure, a ligand and a docking box, and the model returns the ligand's bound 3D pose. It does not find the pocket or model protein flexibility.
Training data come from Binding MOAD. Protein structures were prepared with added hydrogens, protonation states and completed missing heavy atoms and residues, and split 9:1 into training and validation. Fine-tuning ran for 100 epochs on eight V100 GPUs at batch size 64, double the batch of the first version. Input ligand conformers can be supplied or generated from SMILES with standard cheminformatics tools. The released checkpoint is a single 487 MB PyTorch file built on DP Technology's Uni-Core framework.
On the 428-complex PoseBusters V1 set, whose complexes postdate the 2019 cutoff of the training data, 77.6% of ligands were placed within 2.0 Å RMSD and more than 75% of predictions passed all quality checks, against 62.4% for the reprocessed first version, 52.3% for AutoDock Vina and 37.9% for DiffDock. The late-2023 preview of the next AlphaFold generation scored 73.6%. On the 85-complex Astex Diverse set V2 reached 95.29%. Pairing the model with the physics-based Uni-Dock engine raised the share of high-accuracy poses under 1.0 and 1.5 Å.
The model is built for structure-based drug design where the binding site is already known: virtual screening of compound libraries against a target pocket, pose generation for hit follow-up, and producing starting geometries for rescoring or free-energy work. Its authors' case studies span PoseBusters targets including monoacylglycerol lipase, BRAF, HPK1, RORγt and MTH1. A command-line interface accepts PDB, SDF and grid-box JSON files in single, one-to-one batch and one-protein-to-many-ligands modes, and DP Technology hosts it on its Bohrium platform.
Uni-Mol Docking V2 is a technical report rather than a peer-reviewed paper, and its comparisons are re-docking into known pockets, an easier setting than the blind docking many diffusion and co-folding methods attempt. Its main contribution is methodological: showing that much of the gap between ML docking and classical tools on PoseBusters came from input preparation, and that a pretrained encoder can produce poses that pass physical checks without post-hoc repair. The code is released under the MIT license alongside the weights, evaluation sets and the prepared MOAD corpus, and the model has since been adopted as a component elsewhere, including as the docking module of GENzyme.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.