MolTuner: ChEMBL-36 ADMET Property Prediction Dataset with Cluster-Based Train/Test Splits

Multi-task ADMET property prediction dataset filtered from ChEMBL Release 36, used in the MolTuner framework (NeurIPS 2026) for supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT) of small language models. The full ChEMBL_36 corpus (2,239,194 training compounds, 559,755 test compounds) was filtered to small molecules (MW < 1000 Da) that had experimental values for the properties mentioned below. The filtered subset was split into cluster-disjoint train/test sets using BitBirch clustering on RDKit path-based fingerprints (2048 bits). Clusters were assigned to the test set using a distance-from-global-centroid strategy: each cluster's medoid fingerprint was compared to the global centroid of all clusters, and the most structurally peripheral clusters (highest Jaccard distance from the bulk) were greedily selected for the test set until 20% of compounds were reached. Mean top-5 Tanimoto similarity of test compounds to nearest train neighbors is 0.47, confirming a hard scaffold-diverse split. The published files contain the subset of those compounds with at least one measured ADMET property in ChEMBL_36. Each file has 18 property columns (sparse; most molecules have 1-2 measured values): logd, caco2_ab, caco2_ba, mdck_ab, mdck_ba, pampa, pampa_bbb, human_vdss, rat_vdss, mouse_vdss, aqueous_solubility, kinetic_solubility, scaled_clint_hlm, scaled_clint_rlm, scaled_clint_mlm, scaled_clint_hhep, scaled_clint_rhep, scaled_clint_mhep. Select a column and drop NaN rows to recover the single-task subset for any property. Train: 55,552 rows (54,869 unique SMILES), Test: 14,928 rows (14,838 unique SMILES). Fill rate ~8% train, ~6% test. Clustering: BBLean 0.6.1b0, RDKit 2025.9.3, Python 3.11. Source: ChEMBL_36 (https://www.ebi.ac.uk/chembl).