A scaffold split keeps molecules sharing a core chemical structure in the same subset. Training and test compounds therefore come from structurally different scaffold families, making validation more challenging than a random split. section 4.1.1
Randomly separating related molecules can produce overly optimistic results because similar compounds often have similar properties. Scaffold splitting better tests whether a model generalizes to unfamiliar chemical structures. section 4.1.1
A practical scaffold-split workflow is:
- Choose a dataset containing molecular structures and measured targets. section 4.1.1
- Request scaffold-based train, validation, and test partitions rather than random partitions. section 4.1.1
- Compare model performance against random-split results, recognizing that scaffold-based testing is more demanding. section 4.1.1
Listing 4.1 Load AqSolDB with Scaffold-based Splits
from tdc.single_pred import ADME data = ADME(name = 'Solubility_AqSolDB') split = data.get_split(method = 'scaffold') train_df, valid_df, test_df = split['train'], split['valid'], split['test']
Load AqSolDB through TDC and obtain train, validation, and test data using method='scaffold'. section 4.1.1
