macromol-gym 1.1.0

Macromolecular training sets that emphasize protein/non-protein interactions.

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.

[^1]: With the only exceptions being 8h2i and 5zz8. Both of these structures have errors in their biological assemblies, which cause the assemblies to have huge volumes of empty space. Because of the way some internal algorithms work, these particular errors end up requiring prohibitive amounts of memory to process.