jiosephlee/chembl-endpoint-pair-clean-labels

Clean ChEMBL Endpoint Pair Prediction Dataset

This dataset contains rendered question/answer examples from clean ChEMBL endpoint pair records. Answers are sourced from clean aggregated labels: median pChEMBL for continuous endpoints and exact-whitelist 0/1 labels for binary endpoints.

Sampling Procedure

The source is ChEMBL 36. Continuous endpoint labels use non-null pchembl_value rows aggregated by median for each molecule-endpoint. Binary labels use exact whitelist text labels only and exclude rows with pChEMBL values. Binary molecule-endpoint groups with conflicting labels are dropped. Endpoints are retained only when at least two molecules remain after clean-label aggregation.

Endpoint keys are defined at assay scope by (assay_id, standard_type, label_type, binary_semantics) and at target scope by (tid, standard_type, label_type, binary_semantics). The final sample draws one directed molecule pair 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: 353,565 assay endpoints and 33,208 target endpoints
• Final sampled pairs: 100,000 total; 91,413 assay and 8,587 target

Average clean molecules per retained endpoint:

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

split	rows	continuous	binary
train	70,000	44,932	25,068
validation	10,000	6,419	3,581
test	20,000	12,838	7,162

Columns

• question: rendered prompt.
• answer: final answer string, either Answer: <pchembl> or Answer: <0/1>.
• metadata: source identifiers, endpoint metadata, clean labels, and weighted Tanimoto.

Baselines

The baseline suite is included for calibration. copy_reference predicts Molecule B's label by copying Molecule A's clean label. tanimoto_shrinkage blends the copied label toward the train prior using weighted_tanimoto ** beta, with beta tuned on the train split for continuous MAE. For binary rows it also tunes one global decision threshold on the train split. pair_feature_mlp_2layer is a sklearn two-hidden-layer MLP trained separately for continuous regression and binary classification using reference label, pair similarity, endpoint metadata, A/B RDKit descriptor features, and signed B-minus-A descriptor deltas.

Validation Metrics

Continuous

baseline	MAE	RMSE	R2	Pearson	Spearman
copy_reference	0.7480	1.0709	0.4375	0.7186	0.7151
tanimoto_shrinkage	0.7177	0.9875	0.5217	0.7293	0.7235
pair_feature_mlp_2layer	0.7274	0.9648	0.5434	0.7408	0.7338

Binary

baseline	accuracy	balanced accuracy	macro-F1	AUROC
copy_reference	0.9109	0.9103	0.9100	0.9103
tanimoto_shrinkage	0.9106	0.9101	0.9097	0.9360
pair_feature_mlp_2layer	0.9115	0.9111	0.9106	0.9537

Test Metrics

Continuous

baseline	MAE	RMSE	R2	Pearson	Spearman
copy_reference	0.7601	1.0862	0.4152	0.7086	0.7078
tanimoto_shrinkage	0.7291	1.0029	0.5014	0.7177	0.7142
pair_feature_mlp_2layer	0.7272	0.9701	0.5335	0.7334	0.7321

Binary

baseline	accuracy	balanced accuracy	macro-F1	AUROC
copy_reference	0.9038	0.9018	0.9022	0.9018
tanimoto_shrinkage	0.9038	0.9019	0.9022	0.9262
pair_feature_mlp_2layer	0.9024	0.9005	0.9008	0.9387