2023-04-14 Thompson/Wagner Check-in meeting notes

Cases

• Inputs

  • A Topology contains only Molecule

  • A Topology contains only _SimpleMolecule

  • A mix of Molecule and _SimpleMolecule

  • A mix of the above and a new FrozenMolecule subclass that does not yet exist

• Multiple export destinations with different requirements

  • Case where duplication is benign

  • Case where failing to deduplicate is fatal

  • Cases “identical” mols may have different physics params.

  •  

MT ideas

• Have Molecule.are_isomorphic take in 3 or more types (Molecule, _SimpleMolecule, nx.Graph, other FrozenMolecule subclasses)

• Make _SimpleMolecule.are_isomorphic and then when an isomorphism check is needed, look at the types that are to be compared and either call Molecule.are_isomorphic

• Have Molecule.are_isomorphic not know about its inputs, but immediately convert its inputs to graphs via getattr(mol1, "to_networkx")() and error out if that conversion fails.

JW ideas

• Different classes can never be compared (always returns false)

• There's a hierarchy of subclasses where one is always identifiable as "more specific' and will be used for isomorphism checking

• isomorphism checks accept a lambda function, which can change how the comparison is done to allow for use in different contexts

• MT – In favor of “Molecule” compared to “SimpleMol” always fails.

• JW – In favor of “everything turns itself into a networkx graph”, paired “pass in lambda functions describing how you define identity”

• MT – Worth thinking about timelines - I want to get this out ASAP, generalized handling is nice but it’s not clear that people need this soon.

• PAths forward

  • Option 1

    • NOW: Different types are never isomorphic, we implement SimpleMol.are_isomorphioc, objects of the same type can compare to each other

    • LATER: either objects of different types know how to compare to each other, or

  • Option 2

• Long-run decisions

  • Do objects carry around their own comparison logic, or is comparison logic provided externally?

    • If they carry around their own logic, we get into molecule subclass hell, where mols need to know what they’re exporting to to behave right, even before they’ve been exported.

    • So I think the identity-defining logic should not be in the Molecule class

    • To keep with the old API, we need to keep Molecule.are_isomorphic. We could add SimpleMolecule.are_isomorphic but it would be short-lived.

    • Where does the short circuit logic live?

    • Call path could go

      • Topology.custom_grouping(IsomorphismChecker)

      • → short circuit logic

      • → IsomorphismChecker(obj1, obj2)

      • → short circuit logic

      • → Either

        • obj1.is_isomorphic_with(obj2)

        • custom logic, whatever IsomorphismChecker wants.

    • Galaxy brain idea

      • Short circuits are their own objects and know when they’re applicable and when they’re not

  • Short-term:

    • MT want a way to use Topology.identical_molecule_groups to work on a topology that is ONLY _SimpleMolecule

      • This could be implemented by

        • Add _SimpleMolecule.are_isomorphic

        • Inside of the Topology method, compare types:

          • If they are both Molecule, use existing code path

          • If the are both _SimpleMolecule, call the new _SimpleMolecule.are_isomorphic

          • If they are one and the other, return False

        • Inside (simple)Molecule.are_isomorphic

          • return False if other is an unsupported type

          •  

      • Tests involve:

        • Create a Topology of _SimpleMolecule and make sure that Topology.identical_molecule_groups behaves as expected

        • Make sure _SimpleMolecule.are_isomorphic does what we expect

        • Given a mixed topology, make sure checks from either class appropriately short-circuit