Allow kwargs to be passed to force field generators (and then to SMIRNOFF ForceField)

[aclyde11]

It seems forcefield_kwargs is not being propagated sometimes. I've fixed this error below by editing

File "openmmforcefields/generators/template_generators.py", line 1112, in generate_residue_template
    system = self._smirnoff_forcefield.create_openmm_system(molecule.to_topology(), charge_from_molecules=charge_from_molecules)

to

system = self._smirnoff_forcefield.create_openmm_system(molecule.to_topology(), charge_from_molecules=charge_from_molecules, allow_nonintegral_charges = True)

as a hack-y fix.

This is the error I typically get without the fix

openmm_system_generator = SystemGenerator(forcefields=forcefields,
                                                  molecules=[self.mol],
                                                  small_molecule_forcefield=small_molecule_forcefield,
                                                  forcefield_kwargs={'allow_nonintegral_charges' : True})

2020-05-07 14:35:51,119 - openmmforcefields.generators.template_generators - INFO - Requested to generate parameters for residue <Residue 165 (UNL) of chain 1>
2020-05-07 14:35:51,130 - openmmforcefields.generators.template_generators - INFO - Generating a residue template for [H]C1=NN(c2c1c(nc(n2)N([H])[H])N3C(C(C(C(C3([H])[H])([H])[H])([H])C([H])([H])N4C(=C(N=N4)C5(C(C5([H])[H])([H])[H])[H])O[H])([H])[H])([H])[H])[H]
2020-05-07 14:35:51,235 - openforcefield.typing.engines.smirnoff.parameters - INFO - 93 angles added (0 skipped due to constraints)
2020-05-07 14:35:51,427 - openforcefield.typing.engines.smirnoff.parameters - INFO - 51 bonds added (0 skipped due to constraints)
2020-05-07 14:35:51,541 - openforcefield.typing.engines.smirnoff.parameters - INFO - 11 impropers added, each applied in a six-fold trefoil
2020-05-07 14:35:51,917 - openforcefield.typing.engines.smirnoff.parameters - INFO - 144 torsions added
Traceback (most recent call last):
  File "rlmm/tests/test_env.py", line 105, in <module>
    test_load_test_system()
  File "rlmm/tests/test_env.py", line 85, in test_load_test_system
    env.step([0.05, 0.05, 0.05])
  File "/home/aclyde11/RLMM/rlmm/environment/openmmEnv.py", line 72, in step
    self.systemloader.reload_system(gs, test,"test.pdb")
  File "/home/aclyde11/RLMM/rlmm/environment/systemloader.py", line 142, in reload_system
    self.system = openmm_system_generator.create_system(modeller.topology)
  File "/homes/aclyde11/.conda/envs/rlmm/lib/python3.6/site-packages/openmmforcefields/generators/system_generators.py", line 307, in create_system
    system = self.forcefield.createSystem(topology, **forcefield_kwargs)
  File "/homes/aclyde11/.conda/envs/rlmm/lib/python3.6/site-packages/simtk/openmm/app/forcefield.py", line 1148, in createSystem
    templateForResidue = self._matchAllResiduesToTemplates(data, topology, residueTemplates, ignoreExternalBonds)
  File "/homes/aclyde11/.conda/envs/rlmm/lib/python3.6/site-packages/simtk/openmm/app/forcefield.py", line 1352, in _matchAllResiduesToTemplates
    if generator(self, res):
  File "/homes/aclyde11/.conda/envs/rlmm/lib/python3.6/site-packages/openmmforcefields/generators/template_generators.py", line 301, in generator
    ffxml_contents = self.generate_residue_template(molecule)
  File "/homes/aclyde11/.conda/envs/rlmm/lib/python3.6/site-packages/openmmforcefields/generators/template_generators.py", line 1112, in generate_residue_template
    system = self._smirnoff_forcefield.create_openmm_system(molecule.to_topology(), charge_from_molecules=charge_from_molecules)
  File "/homes/aclyde11/.conda/envs/rlmm/lib/python3.6/site-packages/openforcefield/typing/engines/smirnoff/forcefield.py", line 1142, in create_openmm_system
    parameter_handler.postprocess_system(system, topology, **kwargs)
  File "/homes/aclyde11/.conda/envs/rlmm/lib/python3.6/site-packages/openforcefield/typing/engines/smirnoff/parameters.py", line 2786, in postprocess_system
    raise NonintegralMoleculeChargeException(msg)
openforcefield.typing.engines.smirnoff.parameters.NonintegralMoleculeChargeException: Partial charge sum (-2.169999986886978 e) for molecule '_28' (SMILES [H]C1=NN(c2c1c(nc(n2)N([H])[H])N3C(C(C(C(C3([H])[H])([H])[H])([H])C([H])([H])N4C(=C(N=N4)C5(C(C5([H])[H])([H])[H])[H])O[H])([H])[H])([H])[H])[H] does not equal formal charge sum (0). To override this error, provide the 'allow_nonintegral_charges=True' keyword to ForceField.create_openmm_system

[jchodera]

Oh, my! I never considered providing a mechanism to pass arguments to openforcefield.typing.engines.smirnoff.ForceField. I can provide a mechanism to do so, but it will be via a different optional argument to SystemGenerator, perhaps forcefield_generator_kwargs.

Out of curiosity, why would you have non-integral charges?

[aclyde11]

@jchodera why do I have non-integral charges.... ask neural networks. Not sure why that happens. My guess is something gets messed up somewhere. I knew someone would ask when I posted, but I would guess this change would be useful for something relevant as well haha

[jchodera]

You might check out this simple idea in which you ask the ML model to predict two physical parameters (electronegativity and hardness) per atom, and then solve a simple convex optimization problem inspired by a physical charge-equilibration model to determine the (integral) charges. In our hands, this works much better than trying to get the ML model to learn the integral charge constraint!

[ijpulidos]

A mechanism for this has been added for the EspalomaTemplateGenerator in #293 that can be used for the other template generators, but we need to make sure the SMIRNOFFTemplateGenerator know how to handle the parameters given through this mechanism.

[mattwthompson]

296 related, if I understand both accurately