MolTuner: ADMET Property Prediction Dataset from ChEMBL-36

A multi-task dataset of small molecules filtered from ChEMBL Release 36 for predicting ADMET properties. It contains 55,552 training rows and 14,928 test rows, split into structurally distinct clusters to create a hard scaffold-diverse test set. The dataset was created by Aarush Garg for the MolTuner framework and was last updated in May 2026.

Use Cases

• Supervised fine-tuning of small language models for ADMET property prediction based on sparse multi-task data
• Reinforcement fine-tuning of models for molecular property optimization based on the described framework
• Benchmarking model performance on a hard scaffold-diverse split based on the cluster-disjoint train/test design
• Single-task property prediction analysis by selecting and filtering one of the 18 described ADMET property columns

Strengths

• Dataset is explicitly designed for a hard scaffold-diverse split, with a mean top-5 Tanimoto similarity of test compounds to nearest train neighbors of 0.47
• Contains 70,480 total rows (55,552 train, 14,928 test) derived from a large original corpus of 2,239,194 training compounds
• Includes 18 specific ADMET property columns such as logd, caco2_ab, and various scaled clearance metrics

Limitations

• Data is sparse, with fill rates of approximately 8% for train and 6% for test, meaning most molecules have only 1-2 measured values
• Column-level documentation is absent; field semantics must be inferred after download
• License information is unknown, which may restrict usage

Provenance

Source

ChEMBL_36 (https://www.ebi.ac.uk/chembl)

Collection Method

Filtered from ChEMBL Release 36 for small molecules with experimental ADMET values, then split using BitBirch clustering on RDKit fingerprints.

Freshness

Last updated 2026-05-05 01:39:55