Macromolecular Gymnasium

This package builds datasets of macromolecular structures, for machine learning. Some properties of these datasets:

• Capable of processing the entire PDB.¹

• Each data point is a particular coordinate in a particular structure.

• Only data points with sufficient density, and that feature unique combinations of molecules, are included.

• The resulting database is self-contained; all of the necessary coordinates are copied into the database.

• Coordinates are taken from biological assemblies, not asymmetric units.

• Train/validation/test splits are made such that all structures which share an InterPro domain/family end up in the same split.