Should improper torsion RMSDs include all topological torsions or only assigned torsions?

[mattwthompson]

Consider ammonia, which could¹ get assigned improper torsion parameter(s), but conventionally doesn't:

>>> from openff.toolkit import Molecule, ForceField
>>> ammonia = Molecule.from_smiles("N")
>>> ammonia.smirnoff_impropers
{(Atom(name=, atomic number=1), Atom(name=, atomic number=7), Atom(name=, atomic number=1), Atom(name=, atomic number=1)), (Atom(name=, atomic number=1), Atom(name=, atomic number=7), Atom(name=, atomic number=1), Atom(name=, atomic number=1)), (Atom(name=, atomic number=1), Atom(name=, atomic number=7), Atom(name=, atomic number=1), Atom(name=, atomic number=1)), (Atom(name=, atomic number=1), Atom(name=, atomic number=7), Atom(name=, atomic number=1), Atom(name=, atomic number=1)), (Atom(name=, atomic number=1), Atom(name=, atomic number=7), Atom(name=, atomic number=1), Atom(name=, atomic number=1)), (Atom(name=, atomic number=1), Atom(name=, atomic number=7), Atom(name=, atomic number=1), Atom(name=, atomic number=1))}
>>> len(ammonia.smirnoff_impropers)
6
>>> ForceField("openff-2.0.0.offxml").label_molecules(ammonia.to_topology())[0]['ImproperTorsions']
<openff.toolkit.topology.topology.ImproperDict object at 0x17956bbf0>
>>> len([*ForceField("openff-2.0.0.offxml").label_molecules(ammonia.to_topology())[0]['ImproperTorsions']])
0

Should these be accounted for in ICRMSD analysis at all?

Reason to include all topological torsions:

• Significantly easier to implement
• Invariant to different force fields having different SMIRKS patterns
• Might inform (via outliers) good candidates parameter addition?

Reasons to include only assigned torsions

• More intuitive
• Clearer relationship between benchmarking results and force field
• Probably skips many "false positives" (trivalent centers we wouldn't ever assign an improper to)?

---

This is the issue I originally meant to bring up in #199. I used a poor example for that, but surfaced a separate point of confusion so I'm leaving it in place.

Footnotes

1. In the opinion of the toolkit ↩

[lilyminium]

Thanks for the example, Matt. I would vote for it being included as a topological torsion, yes (but not with 6 improper angles -- with 1). My reasoning here is
a) we should know if it ever becomes planar, that's invalid
b) I would want to consistently benchmark the same impropers between force fields even if one of them adds or removes an improper torsion for some reason. Otherwise it's extremely hard to do a like-to-like comparison. For that reason I don't see Clearer relationship between benchmarking results and force field as a straightforward argument against this -- in my ideal benchmarking analysis, we would measure all topological bonds/angles/torsions consistently between force fields, and then do parameter-to-topological comparisons only wrt a single FF (e.g. as is described in the procedure in #190). This avoids issues where re-ordering or modifying parameters means that, e.g., a particular topological angle is in the a31 analysis in Force Field A but in the a32 analysis for Force Field B. (I hope that's clear).

[mattwthompson]

Great, thanks. I'll continue down the topological torsion route, but de-duplicating so that each "6-fold" is only analyzed once