ToxTree: hERG Cardiotoxicity Potency Data for 8879 Compounds

Issar Arab and Khaled Barakat's dataset contains 8879 unique molecular compounds for predicting hERG potassium channel cardiotoxicity liability. The data, gathered from ChEMBL, PubChem, and literature, provides SMILES strings and corresponding pIC50 potency values. It is pre-split into 8380 training and 499 testing compounds for building descriptor-based machine learning models.

Use Cases

• Training binary classification models to predict hERG liability based on molecular structure (SMILES).
• Developing quantitative structure-activity relationship (QSAR) models using pIC50 potency values as a continuous target.
• Benchmarking new machine learning algorithms against a curated, pre-split dataset for cardiotoxicity prediction.
• Validating computational models for off-target cardiac effects in drug discovery pipelines.

Strengths

• Contains 8879 unique molecular compounds, a substantial size for model training.
• Provides explicit pre-defined training (8380) and testing (499) splits for reproducible research.
• Curated from multiple public bioactivity sources (ChEMBL, PubChem) and literature, suggesting data consolidation.

Limitations

• Column names and specific molecular descriptors are not provided in the available metadata.
• The dataset's last update date is unknown, which may affect perceived freshness.
• Metadata completeness is low, lacking information on file format, size, and detailed column descriptions.

Provenance

Source

Compounds gathered from ChEMBL, PubChem, and literature mining.

Collection Method

Built during an NSERC-funded research project for descriptor-based machine learning model development.