PocketEval — HiQBind pocket-level benchmark

PocketEval organises HiQBind at the level of the binding pocket and releases a geometry-aware partition that constrains both pocket-surface similarity and protein-level relatedness.

• 32,230 pocket-level complexes from 18,149 PDB entries
• Released split: 25,784 / 3,223 / 3,223 (train / validation / test)
• Pocket representation: MaSIF-search 80-dimensional surface patch descriptor, mean-pooled to an 80-dimensional pocket vector

Files

File	Size	Content
dataset_index.json	29 MB	one record per pocket: ligand identity, embedding metadata, protein residue count, split
cluster_assignments.json	5.7 MB	cluster_label (MaSIF pocket-similarity connected component) and projection_2d (PCA of the 80-dim descriptor)
benchmark_pocket_split.csv	16 MB	released partition with affinity annotations and grouping columns
summary.json	0.5 kB	dataset-level counts
pocket_structures.tar.gz	1.9 GB	per-pocket pocket_protein.pdb, ligand.sdf, ligand.pdb, metadata.json and the 80-dim embeddings/<pocket_id>.npy
SHA256SUMS.txt	0.5 kB	checksums for the files above

Getting started

# 1. clone the web application (code only)
git clone https://github.com/xyx-2523/PocketEval.git
cd PocketEval

# 2. download the data files from this dataset page into resource_bundle/

# 3. extract the structures into resource_bundle/ (top-level entries are
#    pockets/ and embeddings/, so no --strip-components is needed)
tar -xzf pocket_structures.tar.gz -C resource_bundle/

# 4. verify
sha256sum -c SHA256SUMS.txt

# 5. run the browser
python server.py --host 127.0.0.1 --port 8766
# -> "Pocket web platform ready on http://0.0.0.0:8766 with 32230 pockets"

Once the archive is extracted, pocket search, split and cluster filtering, the 2D embedding projection and the 3D structure viewer all work. /api/structure/<pocket_id>/complex.pdb is composed from pocket_protein.pdb + ligand.pdb at request time, so no separate complex file is distributed.

Layout

resource_bundle/
|-- dataset_index.json
|-- cluster_assignments.json
|-- benchmark_pocket_split.csv
|-- summary.json
|-- pockets/<pocket_id>/{pocket_protein.pdb, ligand.sdf, ligand.pdb, metadata.json}
`-- embeddings/<pocket_id>.npy

Construction

The Pocket partition groups pockets that are similar in MaSIF 80-dimensional surface geometry (mutual patch matching: cosine 0.95, symmetric coverage 0.90, mutual Q25 0.98, support 5) together with pockets sharing a UniProt accession connected component, and assigns each integrated group wholly to one subset. No affinity or ligand information is used to build the graph or the grouping.

Source data

Derived from HiQBind (32,230 complexes; Wang, Y. et al., Digital Discovery 2025, 4, 1209–1220, DOI 10.1039/d4dd00357h), which is distributed under the MIT licence.

Citation

If you use this resource, please cite the PocketEval paper and the HiQBind dataset above.