Data splitters

One important mechanism for avoiding data leakage (see Addressing data leakage in digital reticular chemistry) is the use of proper splitting techniques. mofdscribe implements a range of different approaches, some of which are illustrated in the figure below

As a good default we recommend the use of the HashSplitter, which by default will keep different scaffold sets in different folds.

Note

If you perform a train/test split with a hash splitter and a stratification column, you might be surprised that you will not receive the exact train/test ratio you requested. This is because we perform the stratified split on the group level to be able to guarantee the grouping.