HemanthKari/rcsb-ligand-retrieval-v0.1

The RCSB Protein-Ligand Retrieval Index (v0.1)

An openly licensed dataset of 72,388 protein-ligand complexes for structure-based drug discovery

I built this retrieval index over 72,388 protein-ligand complexes from the RCSB Protein Data Bank so that anyone — academic or commercial — can use it without tripping over licensing. Every component of the release draws on sources that permit free commercial redistribution: RCSB PDB (CC0 1.0), the BindingDB staff-curated affinity subset (CC-BY 3.0), and MIT/BSD-licensed tools (ESM-2, RDKit, FAISS). The release itself is under CC-BY 4.0 — attribution required, nothing else. There is no exposure to academic-only or share-alike datasets like PDBbind, Binding MOAD, ChEMBL, or Matcha.

The primary consumer is the FluxionDock project, where the index serves as the retrieval layer for a retrieval-augmented extension of the MIT-licensed Boltz-1 docking backbone. The index is standalone, though: it is equally useful for protein-ligand similarity search, nearest-neighbor lookup, or as a contrastive pre-training corpus, with no FluxionDock-specific code required.

At a glance

• 72,388 protein-ligand complexes from RCSB PDB
• 1,280-dim pocket embeddings (ESM-2 650M, mean-pooled over binding-site residues within 8 Å of the ligand)
• 2,048-dim ligand embeddings (RDKit Morgan / ECFP4, 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
• 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)	1022 residues

Contents

File	Shape / size	Description
retrieval.faiss	HNSW index	FAISS HNSW32 over joined pocket+ligand embeddings
metadata.jsonl	72,388 rows	pdb_id, ligand resname/chain/resnum, SMILES, pocket_key, offsets
joined_embeddings.npy	(N, 3328) fp32	Unit-normalized [pocket_emb ‖ ligand_emb] for cosine retrieval
pockets/pocket_embeddings.npy	(N, 1280) fp32	ESM-2 mean-pooled embeddings over pocket residues
pockets/pocket_coords.h5	HDF5	Per-complex pocket heavy-atom Cα coordinates
pockets/pocket_manifest.jsonl	72,388 rows	Pocket extraction metadata, residue counts, ligand keys
ligands/ligand_embeddings.npy	(N, 2048) fp32	Morgan fingerprints (radius=2, 2048 bits)
ligands/ligand_coords.h5	HDF5	Per-complex ligand heavy-atom 3D coordinates
ligands/ligand_manifest.jsonl	72,388 rows	SMILES, heavy-atom count, offset into embedding matrix
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	MIT / BSD (depending on backend)	See config.json
Index + metadata JSONL	CC-BY 4.0 (this dataset)	Generated by FluxionDock — 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)

Use cases

The index is a general-purpose retrieval substrate for structure-based drug discovery. A few concrete things it is good for:

1. Retrieval-augmented pose prediction. Condition a diffusion docker (Boltz-1, DiffDock, RFdiffusion-AA) on nearest-neighbor templates instead of sampling from noise. This is the primary FluxionDock use case.
2. Lead hopping and scaffold search. Given a pocket and a SMILES, look up crystallized analogues 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 analogues.
5. Zero-shot ligand similarity. Query by ligand embedding alone 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.

Usage

Full snapshot download

from huggingface_hub import snapshot_download
import faiss, numpy as np, json

local_dir = snapshot_download("HemanthKari/rcsb-ligand-retrieval-v0.1", repo_type="dataset")
index = faiss.read_index(f"{local_dir}/retrieval.faiss")
meta = [json.loads(l) for l in open(f"{local_dir}/metadata.jsonl")]
joined = np.load(f"{local_dir}/joined_embeddings.npy", mmap_mode="r")
print(f"Loaded {len(meta)} complexes, joined shape={joined.shape}")

Nearest-neighbor retrieval by joint query

# Top-5 most similar complexes to the first indexed entry
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(f"  {r['pdb_id']}  lig={r['ligand_resname']:<4}  cos={dist:.4f}")

Pocket-only retrieval (find off-target families)

pocket = np.load(f"{local_dir}/pockets/pocket_embeddings.npy")
pocket = pocket / np.linalg.norm(pocket, axis=1, keepdims=True)  # unit-normalize
# build a small in-memory index over just the pocket slice
import faiss
sub = faiss.IndexFlatIP(pocket.shape[1])
sub.add(pocket)
D, I = sub.search(pocket[[0]], 10)

Ligand-only retrieval (Tanimoto-style scaffold search)

ligand = np.load(f"{local_dir}/ligands/ligand_embeddings.npy")  # (N, 2048) Morgan bits
# Tanimoto similarity on binary fingerprints
query = ligand[0]
inter = (ligand & query.astype(np.uint8)).sum(axis=1)  # for binary
union = (ligand | query.astype(np.uint8)).sum(axis=1)
tanimoto = inter / np.maximum(union, 1)
top10 = np.argsort(-tanimoto)[:10]

Lazy coordinate access via HDF5

import h5py
with h5py.File(f"{local_dir}/pockets/pocket_coords.h5", "r") as f:
    coords = f[f"pockets/{meta[0]['pocket_key']}"][:]  # (n_pocket_residues, 3)
with h5py.File(f"{local_dir}/ligands/ligand_coords.h5", "r") as f:
    lig = f[f"ligands/{meta[0]['pocket_key']}"][:]  # (n_heavy_atoms, 3)

Data schema

Each row in metadata.jsonl has the following fields:

{
  "pdb_id": "1abc",
  "ligand_resname": "HEM",
  "ligand_chain": "A",
  "ligand_resnum": 401,
  "ligand_smiles": "CC1=C(...)",
  "pocket_key": "1abc__HEM_401",
  "pocket_n_residues": 42,
  "offset": 12345
}

The offset field is a direct row index into joined_embeddings.npy, pocket_embeddings.npy, and ligand_embeddings.npy (all three share the same ordering). pocket_key is the HDF5 group name for coordinate lookups.

Provenance

• FluxionDock git commit: 03771f50b0a20a58ecc393957d4cc8bf4800ada7
• Pipeline version: v0.1
• Build timestamp: 2026-04-15T18:51:49Z
• Number of indexed complexes: 72388
• Pocket embedder: ESM-2 (esm2_t33_650M_UR50D, MIT)
• Ligand embedder: RDKit Morgan fingerprints (2048-bit, BSD)
• Upstream RCSB query: resolution <= 2.5 A, has non-polymer ligand, X-ray diffraction

Limitations and known issues

• Resolution cutoff only. I did not filter on R-free, clash score, or Ramachandran outliers. For pose-prediction benchmarks you should add a quality filter downstream.
• One ligand per complex. When a PDB entry contained more than one drug-like ligand, I kept only the first one, so multi-ligand targets are genuinely under-represented.
• Pocket min=1 edge cases. A handful of entries ended up with a single residue within 8 Å of the ligand — tiny ligands or unusual crystal packings. Filter on pocket_n_residues >= 5 if you want larger binding pockets.
• Morgan ligand embeddings. This build ships RDKit Morgan (ECFP4) rather than a 3D-aware neural encoder because the MoLFormer loader path hit a transformers.onnx import error in the build environment. Future versions may ship a Uni-Mol or MoLFormer variant (the ligand_backend field in config.json always records which one was used in a given build).
• No affinity join yet. bindingdb_staff.tsv is shipped as-is rather than pre-joined to the PDB entries. I left it that way so that downstream users can pick their own affinity type (Ki, IC50, Kd) and aggregation policy.

Citation

If you use this dataset, please cite it and the upstream data sources (RCSB PDB, BindingDB, ESM-2) listed in the licensing table. Under CC-BY 4.0 attribution is legally required; in academic practice that means citing this dataset in your paper.

@misc{rcsb_ligand_retrieval_01,
  author = {Hemanth Kari},
  title  = {RCSB Protein-Ligand Retrieval Index (v0.1)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/HemanthKari/rcsb-ligand-retrieval-v0.1}
}