liuganghuggingface/moldata-cara

CARA assay-aware compound activity benchmark v1

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

Part of the Molecular Datasets collection.

Get started

from datasets import load_dataset
dataset = load_dataset("liuganghuggingface/moldata-cara", "lo_all")

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

Tables

Config	Splits	Rows	Targets
lo_all	train, test, unassigned	1,187,136	pChEMBL Value (regression; -log10(M))
vs_all	train, test, unassigned	1,237,256	pChEMBL Value (regression; -log10(M))

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.
• all structures readable: All 2,424,392 SMILES were read with RDKit.
• 19.2% undefined stereo: 19.2% of molecules have a stereocenter whose configuration is not given, so mirror-image forms cannot be told apart.
• 59 stereoisomer sets in a sample: Groups of stereoisomers among 20,000 sampled structures. The dataset is too large to compare all their values.
• 7.1% PAINS alerts: 7.1% of molecules contain a PAINS substructure, which often looks active in screening assays without real activity.
• 896 single-value targets: These targets have the same value for every molecule, so there is nothing to learn from them.
• 30,606 extreme values: 30,606 of 2,420,137 values lie more than 6 robust standard deviations from the median of their target (on a log scale when values span 1,000-fold or more). They are often typing or unit mistakes, and sometimes real; look at them before training.
• 2 repeats over 300× apart: 2 molecules have repeat values more than 300-fold (2.5 log units) apart. In ChEMBL, most gaps this large are entry errors (Kalliokoski et al. 2013).
• mixes AC50, EC50, IC50, Kd…: Column “Value Type” holds AC50, EC50, IC50, Kd, Ki, Potency. Values of different types are not comparable, so train on one type at a time.
• 8 molecules per task: A typical task has 8 labeled molecules. Results on fewer than 1,000 molecules vary a lot between splits.
• source split: train / test: The source divides the data into train (1,868,794 rows), test (323,545 rows), unassigned (232,053 rows). Use this division to compare with published results.
• 2 source splits: The source provides several splits (source, few_shot). The first one gives the training and test sets; each split is also a column named split_, so you can average results over them.

Dataset notes

• Derived from ChEMBL30; overlaps other ChEMBL-based collections.
• Assay identifiers include endpoint type. Distinguish virtual screening from lead-optimization and assay-transfer tasks.
• This is version 1, as used in the paper; Zenodo has a newer version.
• 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.
• The original LO_All test set has 100 assay tasks; VS_All has 96. The source training collection covers additional assays.

License

Dataset-wide license: unspecified. Original source.

Citation

Cite the original study and data source:

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