Project Webpage

A pipeline for curating ChEMBL into a virtual screening dataset for deep-learning model training/validation.

We filter ChEMBL activity data, select artificial decoys with reasonable compound/target criteria. Our pipeline also provides a leakage-resistant train/test splitting by receptor similarity. Using PDBbind+BioLip as default training set, easily replacable with user-custom bulk fasta.

Using just our default setting, you can create the 'ChEMBL-LR' benchmark to test DL models trained on PDBbind and/or BioLip2 dataset wihtout having to worry about target leakage!

All outputs go under a single data directory (e.g. curated_data_filtered/).

conda is the recommended approach - RDKit and nurikit are C++ extension packages that conda resolves cleanly.

conda create -n chemblq python=3.10
conda activate chemblq

# RDKit via conda-forge (simpler than pip for RDKit)
conda install -c conda-forge rdkit

# Remaining Python dependencies
pip install -e .          # installs click, numpy, pandas, requests, tqdm, nurikit

nuri vs nurikit: the Python import is import nuri but the package name on PyPI is nurikit. The pip install -e . above handles this via the dependency in pyproject.toml.

These are not pip/conda packages and must be installed separately:

Note: Structure alignment (Stages 2, 5) uses nurikit (Python TMAlign bindings), which is installed automatically via pip install -e .. No separate TMalign binary is needed.

# Verify MMseqs2 is in PATH
mmseqs --help

DATA=curated_data_filtered

# Stage 1: curate compounds from ChEMBL
chembl-curator curate --download --output $DATA

# Stage 2: validate protein structures and binding sites
chembl-curator filter-proteins --curated-dir $DATA --n-processes 8

# Stage 3: cluster actives per target (Butina, Tanimoto ≥ 0.7)
chembl-curator cluster-actives --data-dir $DATA --workers 8

# Stage 4: build global compound pool
chembl-curator build-pool --data-dir $DATA

# Stage 5: compute pairwise receptor similarity
chembl-curator receptor-sim --data-dir $DATA --mode both --workers 8

# Stage 6: select property-matched, receptor-aware decoys
chembl-curator select-decoys --data-dir $DATA --max-decoys 30

# Stage 7: train/test split by sequence identity clustering
chembl-curator split --data-dir $DATA --valid-frac 0.1

Extracts and filters ligand-target pairs from ChEMBL.

chembl-curator curate --download --output curated_data_filtered
chembl-curator curate --database /path/to/chembl.db --config config.json --output curated_data_filtered
chembl-curator curate --create-config config.json   # generate example config

Default filters:

• Activity types: Ki, Kd, IC50, EC50
• Relations: =, <=
• Units: nM, uM (≤ 10 µM)
• pChEMBL ≥ 5.0, confidence score ≥ 6
• Heavy atoms: 5-80, valid SMILES required
• Binding assays only (B), single protein targets

Configuration (JSON):

{
  "activity_thresholds": {"nM": 10000.0, "uM": 10.0},
  "activity_types": ["Kd", "Ki", "IC50", "EC50"],
  "relations": ["=", "<="],
  "units": ["nM", "uM"],
  "min_pchembl_value": 5.0,
  "min_confidence_score": 6,
  "assay_types": ["B"]
}

Fetches PDB structures, downloads AlphaFold models, aligns structures, and keeps only targets with a single binding site.

chembl-curator filter-proteins --curated-dir curated_data_filtered --n-processes 8

Steps: fetch UniProt PDB list -> download PDB/AlphaFold -> detect ligand-bound structures -> align to AlphaFold (TMalign) -> cluster pockets -> filter single-site targets.

Also writes sequences.fasta (canonical UniProt sequences) and best_structure.tsv (best-resolution ligand-bound structure per target) needed by later stages.

Butina clustering of actives per target. Picks the highest-pChEMBL representative per cluster.

chembl-curator cluster-actives --data-dir curated_data_filtered --dist-thresh 0.3 --workers 8

Builds a global compound pool from all clustered actives. Deduplicates by ChEMBL ID, computes molecular properties and Morgan fingerprints.

chembl-curator build-pool --data-dir curated_data_filtered

Pool contains: MW, cLogP, TPSA, HBD, HBA, aromatic rings, 2048-bit Morgan FP (radius 2), target membership set.

Computes pairwise receptor similarity via two independent methods.

# Both (recommended)
chembl-curator receptor-sim --data-dir curated_data_filtered --mode both --workers 8

# Sequence identity only (faster, no nuri required)
chembl-curator receptor-sim --data-dir curated_data_filtered --mode seqid

Sequence identity: MMseqs2 all-vs-all -> pairwise_seqid.tsv (query, target, seqid 0-1)

Pocket RMSD: TM-align full chain -> filter to pocket residues within radius -> RMSD on ≥3 matched pairs -> pairwise_pocket_rmsd.tsv

Selects property-matched, chemically dissimilar decoys per active. Excludes compounds active against receptors similar to the query target.

chembl-curator select-decoys --data-dir curated_data_filtered --max-decoys 30

Property matching windows: ±50 Da MW, ±2 cLogP, ±50 Ų TPSA, ±2 HBD, ±2 HBA, ±1 aromatic ring.

Train/test split by sequence-identity clustering (MMseqs2). Greedy assignment balances per-source ratios. Per-entry sampling weight = 1 / log2(cluster_size + 1).

A bundled PDBbind+BioLip FASTA (chembl_curator/assets/external_targets.fasta) is included by default.

# Default: includes bundled PDBbind+BioLip sequences
chembl-curator split --data-dir curated_data_filtered --valid-frac 0.1

# ChEMBL-only split (no external sequences)
chembl-curator split --data-dir curated_data_filtered --no-external

# Use your own external FASTA
chembl-curator split --data-dir curated_data_filtered \
    --external-fasta /path/to/your.fasta

External FASTA ID format: IDs must be dot-prefixed as >{source}.{entry_id}, where source is any label (e.g. pdbbind, biolip, myscreendb) and entry_id is any identifier without spaces. For example:

>pdbbind.1a4k
MPPYTVVY...
>biolip.10gs_VWW_A_1
PYTVVYFP...
>myscreendb.custom_entry_42
MKWVTFIS...

Output: train.txt / test.txt (tab-separated, with header):

source	entry_id	compound	weight
chembl	A0A0H2UPP7	CHEMBL405346	0.17
chembl	A0A0H2UPP7	CHEMBL407216	0.17
biolip	10gs_VWW_A_1	-	0.19
pdbbind	1a4k	-	0.26

Output: chembl_targets.tsv (tab-separated, with header):

uniprot	split	n_actives	n_decoys
A0A0H2UPP7	train	2	60
Q9NR56	test	9	270

curated_data_filtered/
├── sequences.fasta             # Canonical sequences, all passed targets
├── best_structure.tsv          # uniprot -> best PDB chain + resolution
├── compound_pool.pkl           # Global compound pool (pickle)
├── pairwise_seqid.tsv          # All-vs-all sequence identity
├── pairwise_pocket_rmsd.tsv    # All-vs-all pocket RMSD
├── passed_targets.txt          # UniProt IDs that passed protein filtering
├── train.txt                   # Train split entries with weights
├── test.txt                    # Test split entries with weights
├── chembl_targets.tsv          # Per-target summary (split, n_actives, n_decoys)
│
└── {UniProt}/                  # Per-target directory
    ├── actives.tsv             # chembl_id, pchembl, smiles
    ├── actives_clustered.tsv   # + cluster_size column
    ├── decoys.tsv              # active_chembl_id -> decoy_ids (;-sep)
    ├── comps/smiles/*.smi      # Raw SMILES files from Stage 1
    ├── pdb/                    # Downloaded PDB + AlphaFold structures
    ├── aligned/                # Structures aligned to AlphaFold model
    ├── pdbid.list              # PDB metadata (method, resolution, chains)
    ├── pocket_info.csv         # Ligand pocket coordinates
    └── sequence.fasta          # Per-target canonical sequence

ChEMBL-Q/
├── chembl_curator/
│   ├── __init__.py
│   ├── cli.py                  # CLI entry points
│   ├── config.py               # CurationConfig
│   ├── curator.py              # Stage 1
│   ├── protein_filter.py       # Stage 2
│   ├── active_clusterer.py     # Stage 3
│   ├── compound_pool.py        # Stage 4
│   ├── receptor_similarity.py  # Stage 5
│   ├── decoy_selector.py       # Stage 6
│   ├── splitter.py             # Stage 7
│   ├── downloader.py
│   ├── filters.py
│   └── assets/
│       ├── excluded_ligands.txt
│       └── external_targets.fasta
├── docs/
│   └── index.html              # Interactive pipeline overview
├── pyproject.toml
└── README.md

• ChEMBL - bioactivity database
• AlphaFold DB - predicted protein structures
• RCSB PDB - experimental protein structures
• MMseqs2 - fast sequence search/clustering
• nurikit - Python TMAlign bindings (used for structure alignment)

This project is provided as-is for research purposes.