liuganghuggingface/moldata-opera_compara

OPERA CoMPARA_QR observed endpoints

Explicitly selected observed endpoint columns from the original OPERA SDF distribution. Other source columns, including model predictions where present, remain metadata.

Processed copy of opera_compara 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://github.com/kmansouri/OPERA
• Original study: OPERA: models for predicting physicochemical properties and environmental fate endpoints
• Built on: US EPA curated data (PHYSPROP and others)

Data quality

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

• info: experiment. Label origin from the source's tags.
• warn: <0.1% unparseable. Rows RDKit cannot parse (kept in quarantine).
• warn: 25% undefined stereo. Molecules with at least one unassigned stereocenter (BACE: 71%).
• warn: 8% PAINS alerts. Molecules matching PAINS substructures, which often interfere with assays.
• warn: 8.5% minority class. Median share of the rarer class across classification tasks.
• bad: 8% label coverage. Median share of molecules labelled per task; missing labels are unmeasured, not negative.
• good: 3,650 molecules per task. Median molecules with a label per task.

Tables

Config	Splits	Rows	Targets
000_CoMPARA_QR	full	47,633	BD_class_Exp_binary (classification, not_verified), AG_class_Exp_binary (classification, not_verified), AN_class_Exp_binary (classification, not_verified)

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("opera_compara")

from datasets import load_dataset
ds = load_dataset("liuganghuggingface/moldata-opera_compara", "000_CoMPARA_QR")

Known issues

• 0 SDF records could not be decoded; original SDF retained.
• Experimental endpoints are explicitly selected; model consensus and predictions are not targets.
• Retain Tr_1_Tst_0 and other source split fields; missing/sentinel values and endpoint units need application-specific checks.
• Repository distribution is a combination of historical releases, not necessarily all OPERA 2.9.5 training updates.

License

Not stated by the source. Redistributed with the collection maintainer's approval; check the original source (https://github.com/kmansouri/OPERA) for its terms before reuse.

Citation

Cite the original study and data source:

• OPERA: models for predicting physicochemical properties and environmental fate endpoints