This repo contains the structure prediction benchmarking dataset from "PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences". (https://doi.org/10.48550/arXiv.2308.05777)
The full set of FASTA files for the 308-structure benchmark set is under sequences/all/. To test on a set that only contains single-chain structures, a 134-structure subset is copied into sequences/monomer/. Finally, a merged FASTA file of all monomers is found in sequences/monomer_combined/
In each directory, there is a no_ligand/ and a ligand subdirectory. The no_ligand/ directory contains just the protein sequences as is typical. To also store ligand information, the ligand/ subdirectory contains FASTA-like files which contain the ligand SMILES as an additional chain. This is the input format for Chai-1.
Similarly, PDB files for the 308-structure benchmark are stored in structures/all/, with the 134-structure monomer subset under structures/monomer/. These files have been stripped of all atoms except for the protein and ligand specified in the original PoseBusters list.
The raw structure files from the protein databank are stored in structures/raw_from_pdb/.
The original list of PDB/ligand IDs is stored in posebuster_benchmark.txt.
Note: 7D6O was replaced with 8J79, as it was a duplicate and marked as obsolete.
To reproduce this dataset, run python scripts/generate_posebusters_benchmark.py in and environment with rdkit and biopython installed.