Index of the 428 recent protein-ligand crystal structures in the PoseBusters docking benchmark, with the 308-complex subset flagged, ligand SMILES, and RCSB links.
PoseBusters paper data (Zenodo), PoseBusters 308-complex identifier list, Buttenschoen, Morris and Deane (2024), Chemical Science, RCSB Protein Data Bank and Chemical Component Dictionary
The PoseBusters Benchmark set is a collection of protein-ligand crystal structures from the Protein Data Bank, assembled by Martin Buttenschoen, Garrett Morris, and Charlotte Deane to test docking methods on complexes they could not have seen in training. Every complex was released in 2021 or later, after the PDBbind General Set v2020 that most deep learning docking models were trained on.
This table indexes all 428 complexes in the original release and flags the 308-complex subset used in the journal paper. Each row pairs the PDB entry with its ligand code, the ligand structure as SMILES, and the protein it binds, so you can look up a target, check its resolution, and load it into a docking tool.
How the set was assembled
The authors selected recent, high-quality crystal complexes containing drug-like ligands; the 308-complex set used in the paper has 308 unique proteins and 308 unique ligands. Each complex was downloaded from the PDB and processed with PyMOL to remove solvent and every copy of the ligand of interest, leaving the protein with its cofactors as a PDB file and the ligand as an SDF file.
During peer review the authors found that some of the 428 structures contain crystal contacts, so the results in the journal paper use a subset of 308 complexes. Later papers often call these two versions PoseBusters V1 and V2. Use the "In 308 set" column to filter for the subset.
What counts as a successful prediction
The usual pose accuracy measure is the symmetry-aware heavy-atom RMSD between the predicted ligand and the crystal ligand, with a prediction counted as correct when it falls within 2 Å. PoseBusters adds a second requirement: the pose must also be physically and chemically plausible. Its checks cover RDKit sanitization, unchanged molecular formula, bonds and stereochemistry, bond lengths and angles, flat aromatic rings and double bonds, internal clashes, a conformer energy ratio, and clashes or volume overlap with the protein and cofactors. A pose that passes all of them is called PB-valid.
In the original paper, AutoDock Vina and Gold were the strongest methods on this set. The best deep learning method, DiffDock, succeeded on about 12% of complexes, and the deep learning methods did worst on proteins with low sequence identity to their training data. The sequence identity column lets you reproduce that split.
Using the set in ProteinIQ
The send button on each row loads the RCSB entry and the ligand SMILES into a docking tool such as AutoDock Vina, GNINA, or DiffDock-L. The RCSB file is the full deposited entry, so it still contains the crystal ligand, waters, and any other chains. Remove the ligand before docking, or use the prepared protein files from Zenodo, which already have it removed.
To score your poses the way the paper did, run them through PoseBusters on ProteinIQ with the predicted ligand, the crystal ligand, and the protein. It reports each check separately along with the RMSD, so you can see why a pose fails rather than only whether it does.
How this dataset was assembled
Complex membership, sequence identity, and cofactor flags come from the crystal-structure rows of the PoseBusters paper results on Zenodo. The 308-complex subset is the identifier list the authors published alongside it. Release date, resolution, and the protein chain nearest each ligand come from the RCSB PDB Data API, and ligand names and SMILES from the Chemical Component Dictionary. Every SMILES string parses in RDKit, which also supplies the formula and molecular weight.
Each row keeps the PDB ID and ligand code that identify the complex in the original benchmark, so any entry can be checked against RCSB or the Zenodo files before reuse. One entry, 7D6O, has since been removed from the PDB and is kept here because it is part of the published set.
License and citation
This compilation is free to download and reuse under the Creative Commons Attribution 4.0 International license. Credit ProteinIQ and the original sources when you share it or publish results that use it.
Suggested citation: ProteinIQ (2026). PoseBusters Benchmark set, version 2026.10.05. https://proteiniq.io/datasets/posebusters-benchmark
ProteinIQ does not host the prepared protein and ligand files. Download them from the PoseBusters Zenodo record, which is released under CC BY 4.0.
Sources and further reading
PoseBusters paper data (Zenodo): Complex membership, per-complex sequence identity, cofactor flags, and the prepared protein and ligand files. Released under CC BY 4.0.
RCSB Protein Data Bank and Chemical Component Dictionary: Release dates, resolution, the protein chain nearest each ligand with its UniProt accession, and ligand names and SMILES from the Chemical Component Dictionary. PDB data are released under CC0 1.0.
Frequently asked questions
The original release has 428 complexes. The journal paper reports results on a 308-complex subset after the authors found crystal contacts in some structures. Both are in this table, with the subset flagged.
V1 usually refers to the 428-complex release and V2 to the 308-complex subset used for the journal results. Check which one a paper used before comparing numbers.
A pose is PB-valid when it passes every PoseBusters check for chemical consistency, ligand geometry, and clashes with the protein and cofactors. Benchmarks usually report the share of poses that are both PB-valid and within 2 Å RMSD of the crystal ligand.
No. This page offers an index of the complexes as CSV or JSON. The prepared protein and ligand files are on the PoseBusters Zenodo record, and each row links to its RCSB entry.
Cite Buttenschoen, Morris and Deane, "PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences", Chemical Science 15, 3130 (2024), doi:10.1039/D3SC04185A, and the Zenodo record at zenodo.org/records/8278563.