May ultimately need different handling of long range charge interactions

[davidlmobley]

In the context of SAMPL6 host-guest binding, Andrea Rizzi noted they had run into this issue:

Since it’s not possible in OpenMM’s custom forces to compute the reciprocal space component of PME, in YANK, we simulate using only the direct space, and we correct for this by reweighting the end states to a non-alchemical system using MBAR. This has worked very well so far, but, for several ligands in the SAMPL6 set, I’ve noticed that the overlap between the alchemical and non-alchemical states is poor and we are getting high variability among the SAMPLing replicates (in a couple of cases on the order of 10kcal/mol).

I believe this is what you [and @elkhoury ] were observing as well with charged ligands a couple of months ago, although back then we mistook it for a problem caused by long correlation times associated to the counterions sampling, and I was hoping that running long simulations would avoid this problem.

They are planning to correct the issue.

John points this out:

The current plan is to add support for changing the charges directly, but that won't be efficient for BLUES. I think it is worthwhile to revisit in the OpenMM issue tracker how we can either introduce Context parameters that scale whole Forces or force groups, or how we can implement a more sophisticated CHARMM BLOCk-like capability

He suggests looking at https://github.com/pandegroup/openmm/issues/1702 and how they might implement a solution, https://github.com/choderalab/openmmtools/issues/319 .

We should take a look.

[davidlmobley]

It's possible this also could affect acceptance rates for relatively more polar and especially charged ligands in BLUES; while it's reasonable to expect that these might be lower, perhaps an issue of this type makes them even lower than they should be (impacting the results of @nathanmlim ).

[nathanmlim]

The option to treat this is provided here

With the parameter alchemical_pme_treatment, I'm not exactly sure what the best option to use here is but it's been set to the same default in openmmmtools:

        alchemical_pme_treatment : str, optional, default = 'direct-space'
            Controls how alchemical region electrostatics are treated when PME is used.
            Options are ['direct-space', 'coulomb', 'exact'].
            - 'direct-space' only models the direct space contribution
            - 'coulomb' includes switched Coulomb interaction
            - 'exact' includes also the reciprocal space contribution, but it's
                only possible to annihilate the charges and the softcore parameters
                controlling the electrostatics are deactivated. Also, with this
                method, modifying the global variable `lambda_electrostatics` is
                not sufficient to control the charges. The recommended way to change
                them is through the `AlchemicalState` class.