jiosephlee/chembl-endpoint-context10-clean-labels

Clean ChEMBL Endpoint Context-10 Prediction Dataset

This dataset contains rendered question/answer examples from clean ChEMBL endpoint records. Each example gives 10 labeled reference molecules from one endpoint and asks for the clean label of one held-out query molecule from the same endpoint.

Sampling Procedure

The source is ChEMBL 36. Continuous labels are median pchembl_value per molecule-endpoint. Binary labels are exact whitelist text labels only; rows with pChEMBL are excluded and molecule-endpoint conflicts are dropped. Endpoints are retained for this version only when at least 11 clean labeled molecules remain, so each row can contain 10 references plus one query.

Endpoint keys are assay-scope (assay_id, standard_type, label_type, binary_semantics) or target-scope (tid, standard_type, label_type, binary_semantics). The final sample draws one context-10 example per endpoint, proportionally across endpoint scope, label type, and binary semantics.

Pool Statistics

• Clean endpoint source rows: 10,154,459
• Binary whitelist rows: 5,257,668
• Binary conflict groups dropped: 48,597
• Clean endpoint pool before >=11 filter: 386,773 total; 353,565 assay and 33,208 target
• Clean endpoint pool after >=11 filter: 115,813 total; 99,351 assay and 16,462 target
• Final sampled rows: 99,960 total; 85,747 assay and 14,213 target

Average clean molecules per retained endpoint before the >=11 filter:

endpoint group	avg molecules
assay overall	27.33
target overall	253.20
assay continuous pChEMBL	20.17
target continuous pChEMBL	234.50
assay binary activity	47.32
target binary activity	289.62

Splits

The split is stratified by metadata.label_type using 70/10/20 train/validation/test.

split	rows	continuous	binary
train	69,972	55,964	14,008
validation	9,996	7,995	2,001
test	19,992	15,990	4,002

Columns

• question: rendered prompt with 10 labeled reference molecules and one unlabeled query molecule.
• answer: final answer string, either Answer: <pchembl> or Answer: <0/1>.
• metadata: endpoint identifiers, query label, reference labels, and query-reference weighted Tanimoto values.

Baselines

The baseline suite uses only the 10 clean reference labels and query-reference weighted Tanimoto values.

Validation Metrics

Continuous

baseline	MAE	RMSE	R2	Pearson	Spearman
reference_mean	0.5910	0.8092	0.6689	0.8186	0.8147
reference_median	0.5920	0.8383	0.6447	0.8082	0.8050
similarity_weighted_mean	0.5643	0.7776	0.6943	0.8337	0.8303
nearest_reference_copy	0.6055	0.8830	0.6058	0.8023	0.7995

Binary

baseline	accuracy	balanced accuracy	macro-F1	AUROC
reference_majority	0.9105	0.8754	0.8783	0.9612
reference_positive_rate	0.9105	0.8754	0.8783	0.9612
similarity_weighted_vote	0.9280	0.8932	0.9009	0.9675
nearest_reference_copy	0.9250	0.8967	0.8984	0.8967

Test Metrics

Continuous

baseline	MAE	RMSE	R2	Pearson	Spearman
reference_mean	0.5791	0.7879	0.6733	0.8214	0.8206
reference_median	0.5791	0.8151	0.6503	0.8119	0.8116
similarity_weighted_mean	0.5507	0.7544	0.7005	0.8375	0.8371
nearest_reference_copy	0.6011	0.8694	0.6021	0.8018	0.8009

Binary

baseline	accuracy	balanced accuracy	macro-F1	AUROC
reference_majority	0.9150	0.8867	0.8859	0.9650
reference_positive_rate	0.9150	0.8867	0.8859	0.9650
similarity_weighted_vote	0.9293	0.8992	0.9038	0.9690
nearest_reference_copy	0.9203	0.8956	0.8934	0.8956