liuganghuggingface/moldata-pcba_128

PubChem BioAssay 128-assay molecular screening benchmark

PubChem BioAssay 128-assay molecular screening benchmark. DeepChem/MoleculeNet release.

Part of the Molecular Datasets collection.

Get started

from datasets import load_dataset
dataset = load_dataset("liuganghuggingface/moldata-pcba_128", "000_pcba")

Where the data comes from

• Original data: https://pubchem.ncbi.nlm.nih.gov/
• Redistributed by: MoleculeNet (DeepChem) (dataset page)

Tables

Config	Splits	Rows	Targets
000_pcba	full	437,927	PCBA-1030 (classification; binary), PCBA-1379 (classification; binary), PCBA-1452 (classification; binary), PCBA-1454 (classification; binary), PCBA-1457 (classification; binary), PCBA-1458 (classification; binary), PCBA-1460 (classification; binary), PCBA-1461 (classification; binary) and 120 more

data/ contains rows with readable molecular structures; identical observations are collapsed. observations/ contains all source rows.

Source label columns are text, including values such as <10. SMILES_original retains the input representation; SMILES_canonical provides a canonical representation. structure_key groups molecular components, and structure_status records parsing outcomes.

Data quality

• experiment: Labels are measured in a lab.
• 0% unreadable: 2 rows of 437,929 have a SMILES that RDKit cannot read; they are left out of the processed table.
• 25.7% undefined stereo: 25.7% of molecules have a stereocenter whose configuration is not given, so mirror-image forms cannot be told apart.
• 2,068 of 222,558 stereoisomer sets differ: Found 222,558 groups of molecules that differ only in stereochemistry. In 2,068, the values are 10-fold apart (log-scale labels), more than one standard deviation apart, or in different classes. That can be a real effect or a mistake in the source; a model that ignores stereochemistry treats these molecules as the same.
• 5.2% PAINS alerts: 5.2% of molecules contain a PAINS substructure, which often looks active in screening assays without real activity.
• 0.5% in rarer class: In a typical task, 0.5% of labeled molecules are in the rarer class. Accuracy is misleading here; use precision and recall.
• 70.6% measured per task: In a typical task, 70.6% of molecules have a value. A missing value means not measured, not inactive.
• 309,587 molecules per task: A typical task has 309,587 labeled molecules.
• no source split: Make your own split. A random split puts near-identical molecules in both training and test sets, so results look better than they will on new compounds. Splitting by scaffold, or by date when dates are given, gives a more realistic estimate.

Dataset notes

• Historical PubChem assay snapshot. Missing labels mean not measured. Classes are highly imbalanced, some actives may be assay interference, and compounds overlap other PubChem benchmarks.

License

Dataset-wide license: unspecified. Original source.

Citation

Cite the original study and data source:

• PubChem BioAssay data as curated for MoleculeNet
• Wu et al. MoleculeNet: a benchmark for molecular machine learning (2018)
• Official DeepChem loader and task definitions