pytdc

You can load standardized, AI-ready ADME prediction datasets using predefined benchmark splits. It provides train, validation, and test splits, including scaffold splits, for reproducible property prediction modeling.

To evaluate drug-target interaction models, you can use cold splits that separate unseen drugs or targets. This ensures realistic benchmarking for interaction prediction tasks across standardized therapeutic datasets.

Yes, molecule generation and retrosynthesis workflows are supported. You can access goal-directed generation datasets and use provided script templates to benchmark oracle evaluation for therapeutic drug discovery tasks.

Yes, PyTDC installation is required. Once installed, you can use the get_split and get_data interfaces to retrieve standardized toxicity prediction datasets and apply them to your modeling pipelines.

Scaffold splitting groups molecules by their core chemical structures into separate train and test sets. This prevents data leakage in single-instance prediction tasks like ADME, toxicity, and quantum mechanics modeling.

Pre-defined benchmark splits enforce specific data partitioning protocols for drug-drug interaction evaluation. If your research requires custom data splitting logic, you may need to manually configure splits outside the provided interfaces.