Overview of Dataset Splits

Data split functions partition a dataset into training, validation, and test sets to support training of ML models, hyperparameter tuning, and fair evaluation and comparison of ML models. The function is called from within the data loader class. It accepts the following three inputs:

  • method: Type of dataset split. TDC provides several types dataset splits to support realistic evaluations. Default type is Random Split.
  • seed: random seed.
  • frac: proportional size of training, validation, and test sets. By default, the sizes are set to 70%, 10%, and 20%, respectively.

As the default TDC data format is Pandas DataFrame, it will return a dictionary with key 'train', 'valid', and 'test' and value of each set's data frame.

from tdc.X import Y
data = Y(name = Z)
split = data.get_split(method = 'random', seed = 42, frac = [0.7, 0.1, 0.2])
# split: {'train': train dataframe, 'valid': valid dataframe, 'test': test dataframe}

Important Note: in this part, TDC provides a generic data split function where you can tune splitting schemes for your various research needs. In the benchmark mode, it is set to be a specific splitting method with fixed seed and fractions. To get that split, please use the BenchmarkGroup class in the Leaderboard page.

We continue with brief description of different kinds of dataset splits provided in the TDC

Random Split

Description: The default for any split function. Randomly split the data into train, validation, and test set.

from tdc.single_pred import ADME
data = ADME(name = 'Caco2_Wang')
split = data.get_split(method = 'random')

Scaffold Split

Description: Scaffold split is based on the scaffold of the molecules so that train/val/test set is more structurally different. It is more challenging than random split.

Note: Scaffold split only applies to single-instance drugs-related tasks (ADME, Tox, HTS). Scaffold split also requires RDKit installed.

from tdc.single_pred import ADME
data = ADME(name = 'Caco2_Wang')
split = data.get_split(method = 'scaffold')

Cold-Start Split

Description: Cold-start split is for multi-instance prediction problems such as DTI, GDA, DrugRes, MTI, where they present two entity types. It first splits on one entity type into train/valid/test and then move all pairs associated with that entity in each set as the final splits. To use that, set column_name to be the entity that you want to split on. For example, to do cold drug split on DTI task:

from tdc.multi_pred import DTI
data = DTI(name = 'DAVIS')
split = data.get_split(method = 'cold_split', column_name = 'Drug')

Combination Split

Description: Drug Combination Split is for drug combination dataset, where we split on drug combinations such that training/validation/testing would have distinct set of combinations.

Note: Combination split only applies to drug combination tasks such as DrugSyn.

from tdc.multi_pred import DrugSyn
data = DrugSyn(name = 'DrugComb')
split = data.get_split(method = 'combination')