PubChem as a dynamic dataset

[sfluegel05]

This originated from the March 26 StrOntEx meeting.

Status Quo

The ChEBI datasets all extend the _DynamicDataset class. This means that they produce

• a data.pkl file with all molecule-label pairs in ChEBI
• a splits.csv file matching ChEBI IDs to train/val/test subsets

This has the advantage of enhancing reproducibility (e.g. by only supplying the splits file and chebai version used)

Goal

Implement the same for PubChem