liuganghuggingface/moldata-cara

CARA assay-aware compound activity benchmark v1

Original CARA v1 archive, assay-aware activity data and benchmark split definitions.

Processed copy of cara from the molecular datasets collection. Structures are canonicalized and validated; labels are never averaged or filled. Prefer the original source for citation.

Where the data comes from

• Original data: https://zenodo.org/records/11063965
• Original study: Benchmarking compound activity prediction for real-world drug discovery applications
• Built on: ChEMBL

Data quality

Grade: care (rough; rules on the hub page).

• info: experiment. Label origin from the source's tags.
• good: valid structures. Rows RDKit cannot parse (kept in quarantine).
• warn: 19% undefined stereo. Molecules with at least one unassigned stereocenter (BACE: 71%).
• info: 59 stereoisomer sets. Groups of stereoisomers in the sampled structures; their labels can differ sharply.
• warn: 7% PAINS alerts. Molecules matching PAINS substructures, which often interfere with assays.
• bad: 8 molecules per task. Median molecules with a label per task.

Tables

Config	Splits	Rows	Targets
000_LO_All.tsv	full	1,187,136	pChEMBL Value (regression, negative log10 molar activity; Value Type identifies experimental endpoint)
001_VS_All.tsv	full	1,237,256	pChEMBL Value (regression, negative log10 molar activity; Value Type identifies experimental endpoint)

data/ holds parseable rows with identical rows collapsed. observations/ holds every source row, including unparseable structures, for auditing. Added columns: SMILES_original, SMILES_canonical (RDKit; stereo, salts, and * kept), structure_key (JSON list of canonical components; use it to deduplicate or group), structure_status, wildcard_atom_count, source_row_first, observation_count, source_dataset, source_table. All original columns are strings, so censored values such as <10 survive; cast labels yourself.

Load

from torch_molecule.datasets import load_dataset
data = load_dataset("cara")

from datasets import load_dataset
ds = load_dataset("liuganghuggingface/moldata-cara", "000_LO_All.tsv")

Known issues

• Derived from ChEMBL30; overlaps other ChEMBL-based collections.
• Assay identifiers include endpoint type. Distinguish virtual screening from lead-optimization and assay-transfer tasks.
• Version pinned to original v1 paper record; Zenodo indicates a newer version exists.
• GPCR/Kinase files are subset views of LO_All/VS_All and are not additional independent observations. LO/VS task scenarios can share source activity observations.
• Source assay coverage is much larger than held-out benchmark coverage. Original LO_All test JSON has100 assay task IDs; VS_All test JSON has96. Do not report all training/source assays as held-out benchmarks.

License

Not stated by the source. Redistributed with the collection maintainer's approval; check the original source (https://zenodo.org/records/11063965) for its terms before reuse.

Citation

Cite the original study and data source:

• Benchmarking compound activity prediction for real-world drug discovery applications
• Original CARA implementation