HemanthKari/rcsb-ligand-retrieval-v0.1

The RCSB Protein-Ligand Retrieval Index

An openly licensed retrieval dataset of 72,199 protein-ligand complexes from the RCSB Protein Data Bank, built for retrieval-augmented structure-based drug discovery. Every component of the release is commercially redistributable end-to-end.

Version history

Date	Change
2026-04-15	Initial release: FAISS HNSW index, ESM-2 pocket embeddings, Morgan ligand embeddings, HDF5 coordinate files, BindingDB staff-curated affinities.
2026-04-18	Added protein_seq field to metadata (sequences fetched from RCSB FASTA endpoint, longest protein chain per entry; 71,742 / 72,199 rows populated).
2026-05-11	Added canonical splits/train_indices.npy and splits/val_indices.npy (seed=42, val_frac=0.05); 68,459 train / 3,608 val entries filtered for usable HDF5 ligand + pocket coordinates.
2026-05-01	Correction. Regenerated joined_embeddings.npy, retrieval.faiss, and metadata.jsonl after fixing a join bug in the index builder: the previous version paired pocket and ligand embeddings by row position rather than by pocket_key, causing the ligand portion of the joined embedding to be misaligned by 1–N positions for entries past the first ligand-embedding failure. Per-complex pocket_embeddings.npy, ligand_embeddings.npy, the HDF5 coordinate files, and the splits are unaffected and were not re-uploaded.
2026-05-01	Hygiene. The joined retrieval index now excludes 189 entries whose RDKit Morgan fingerprint returned a zero-vector (transition-metal complexes, polyborane clusters, polyoxometalates, and other ligands with no recognizable organic substructure). Joined-index row count: 72,388 → 72,199. Per-complex ligand_embeddings.npy and ligand_coords.h5 retain the original 72,388 entries; only the joined metadata.jsonl, metadata.parquet, joined_embeddings.npy, retrieval.faiss, and the canonical splits were updated.

At a glance

• 72,199 protein-ligand complexes in the joined retrieval index (per-complex pocket and ligand artifacts cover the full 72,388 pre-filter entries; see Contents)
• 1,280-dim pocket embeddings (ESM-2 650M, mean-pooled over binding-site residues within 8 Å of the ligand)
• 2,048-bit Morgan / ECFP4 ligand fingerprints (radius 2)
• 3,328-dim joined index (unit-normalized, cosine-ready)
• FAISS HNSW index (M=32, efConstruction=200) for sub-millisecond nearest-neighbor lookup
• BindingDB staff-curated affinities (CC BY 3.0) for supervised training signals
• Protein sequences (FASTA, longest chain per PDB)
• Canonical train/val splits for reproducible benchmarking
• CC BY 4.0 — free commercial use with attribution

Filter criteria

The underlying RCSB query and drug-like filter:

Criterion	Value
Experimental method	X-ray diffraction
Max resolution	2.5 Å
Min protein length	50 residues
Ligand heavy atoms	10–70 (drug-like range)
Pocket radius	8.0 Å around ligand heavy atoms
Max sequence length (for ESM-2)	1,022 residues

Contents

File	Shape / size	Description
retrieval.faiss	HNSW index	FAISS HNSW index (M=32, efConstruction=200) over joined pocket+ligand embeddings (72,199 entries)
metadata.jsonl	72,199 rows	pdb_id, ligand resname/chain/resnum, SMILES, pocket_key, offsets, protein_seq
metadata.parquet	72,199 rows	Same as metadata.jsonl in columnar form (used by the HF viewer)
joined_embeddings.npy	(72,199, 3328) fp32	Unit-normalized [pocket_emb ‖ ligand_emb] for cosine retrieval
pockets/pocket_embeddings.npy	(72,485, 1280) fp32	ESM-2 mean-pooled embeddings over pocket residues
pockets/pocket_coords.h5	72,485 datasets	Per-complex pocket residue heavy-atom mean coordinates
pockets/pocket_manifest.jsonl	72,485 rows	Pocket extraction metadata, residue counts, ligand keys
ligands/ligand_embeddings.npy	(72,388, 2048) fp32	Morgan fingerprints (radius=2, 2048 bits)
ligands/ligand_coords.h5	72,388 datasets	Per-complex ligand heavy-atom 3D coordinates (world frame, PDB convention)
ligands/ligand_manifest.jsonl	72,388 rows	SMILES, heavy-atom count, offset into embedding matrix
splits/train_indices.npy	(N_train,) int64	Metadata row indices for the canonical training split
splits/val_indices.npy	(N_val,) int64	Metadata row indices for the canonical validation split
splits/splits_info.json	JSON	Split parameters (seed, val_frac, counts)
raft_pdb_filtered.jsonl	Manifest	All complexes passing the drug-like filter (pre-embedding)
bindingdb_staff.tsv	~93K rows	BindingDB staff-curated affinity subset (CC BY 3.0)
config.json	JSON	Exact filter/embedding parameters, git SHA, build timestamp

Commercial licensing

Every component of this dataset is commercially redistributable:

Component	License	Source
Protein-ligand 3D coordinates	CC0 1.0 Public Domain	RCSB Protein Data Bank
Binding affinity labels	CC BY 3.0	BindingDB staff-curated subset
ESM-2 pocket embeddings	MIT (via ESM-2 model weights)	facebookresearch/esm
Ligand embeddings	BSD	RDKit Morgan fingerprints
Index + metadata JSONL	CC BY 4.0 (this dataset)	Generated by this pipeline; attribution required

Explicitly excluded (not used in any stage of this dataset):

• PDBbind (any version) — academic or paid
• Binding MOAD — CC BY-NC 4.0
• ChEMBL — CC BY-SA 3.0 share-alike
• Matcha — CC BY-NC 4.0
• CASF-2016
• Astex Diverse Set raw CCDC release (use PoseBench CC BY 4.0 preprocessed version instead)

Usage

Full snapshot download

from huggingface_hub import snapshot_download
import faiss
import json

local_dir = snapshot_download("HemanthKari/rcsb-ligand-retrieval-v0.1", repo_type="dataset")
index = faiss.read_index(local_dir + "/retrieval.faiss")
with open(local_dir + "/metadata.jsonl") as f:
    meta = [json.loads(l) for l in f]
print("Loaded %d complexes" % len(meta))

Nearest-neighbor retrieval

import numpy as np

joined = np.load(local_dir + "/joined_embeddings.npy", mmap_mode="r")
q = joined[0].astype("float32").reshape(1, -1)
D, I = index.search(q, 5)
for dist, idx in zip(D[0], I[0]):
    r = meta[idx]
    print("  %s  lig=%-4s  cos=%.4f" % (r["pdb_id"], r["ligand_resname"], dist))

Pocket + ligand coordinates via HDF5

import h5py

with h5py.File(local_dir + "/pockets/pocket_coords.h5", "r") as f:
    pocket_xyz = f["pockets/" + meta[0]["pocket_key"]][:]  # (N_pocket_residues, 3) — per-residue mean heavy-atom positions
with h5py.File(local_dir + "/ligands/ligand_coords.h5", "r") as f:
    ligand_xyz = f["ligands/" + meta[0]["pocket_key"]][:]  # (N_heavy_atoms, 3)

Coordinates are stored in the deposited Protein Data Bank reference frame. Consumers who require a per-complex canonical reference frame can construct one through principal-component analysis of the pocket alpha-carbon coordinates with a deterministic ligand-oriented sign convention.

Canonical train/val split

import numpy as np

train_ix = np.load(local_dir + "/splits/train_indices.npy")
val_ix = np.load(local_dir + "/splits/val_indices.npy")
print("train=%d, val=%d" % (len(train_ix), len(val_ix)))

# Each entry is an index into metadata.jsonl / joined_embeddings.npy.
train_meta = [meta[i] for i in train_ix]

Use cases

The index is a general-purpose retrieval substrate for structure-based drug discovery. Concrete applications include:

1. Retrieval-augmented pose prediction. Condition a diffusion docker (Boltz-1, DiffDock, RFdiffusion-AA) on nearest-neighbor templates instead of sampling from noise.
2. Lead hopping and scaffold search. Given a pocket and a SMILES, look up crystallized analogs with similar binding environments.
3. Virtual screening benchmark construction. Draw time-split or structure-split negative controls from the index without touching academic-only sets like DUD-E or PDBbind.
4. Pocket similarity search. Query by pocket embedding alone to surface off-target families or allosteric-site analogs.
5. Zero-shot ligand similarity. Query by Morgan ligand embedding for ECFP-based Tanimoto-style retrieval across 72K crystallized drug-like molecules.
6. Dataset bootstrapping for embedding models. Train contrastive or triplet-loss models for pocket-ligand matching without having to renegotiate a license for every new experiment.
7. Structural biology QA / RAG. Ground an LLM's answers about specific binding sites with real PDB coordinates.

Citation

If you use this dataset, please cite it and the upstream data sources (RCSB PDB, BindingDB, ESM-2) listed in the licensing table.

The dataset has a persistent identifier: DOI: 10.57967/hf/8408.

@misc{kari2026rcsb,
  author       = {Hemanth Kari},
  title        = {Pocket and ligand embeddings for 72,199
                   protein-ligand complexes from the Protein Data Bank},
  year         = {2026},
  howpublished = {Hugging Face},
  doi          = {10.57967/hf/8408},
  url          = {https://doi.org/10.57967/hf/8408},
  note         = {Released under CC BY 4.0}
}

Pipeline source code

The pipeline source code that constructs this dataset is deposited on Zenodo under the MIT license as a versioned, immutable archive:

Pipeline source DOI: 10.5281/zenodo.19954940

The dataset itself, as released here, is independent of any downstream model that may consume it and imposes no licensing back-obligation on consumers. The exact pipeline commit identifier active during the v0.1 build is recorded in config.json.