pytdc

You can load standardized therapeutics datasets for machine learning evaluation using a unified API that provides task-scoped access to ADME, toxicity, DTI, DDI, and molecular generation benchmarks. This ensures consistent formats and reproducible experimental protocols.

A scaffold split partitions chemical data based on structural diversity, ensuring test molecules have distinct scaffolds from training data. You need it for molecular property prediction to evaluate model generalization on novel chemical structures rather than memorized analogs.

Yes, you can generate cold splits for DTI and DDI generalization scenarios. The system supports cold-start splits that hold out unseen drugs and/or targets, allowing you to rigorously evaluate interaction prediction models on novel entities.

The best way to benchmark molecule optimization workflows is using built-in molecular property oracles. These oracles score candidate molecules for goal-directed optimization, providing standardized evaluation metrics for property prediction and generation tasks.

Yes, the therapeutics benchmarking tool supports retrosynthesis tasks. It provides curated datasets and standardized evaluators for molecule generation and retrosynthesis, ensuring AI-ready data formats for machine learning experimentation.

To evaluate drug discovery models consistently across multiple seeds, apply the standardized evaluator workflow. It enables benchmark-group evaluation across multiple random seeds, yielding robust performance comparisons and highly reproducible results.