Re: [AMBER] QMMM free energy calculations for alchemical changes using a coupling parameter in amber 18

From: E. Kuechler <erich.kuechler.gmail.com>
Date: Thu, 29 Nov 2018 17:13:32 -0800

At least throw me a bone, Brian ;)

I have done a study in which I performed the MM->QM/MM correction TI that's
being discussed, here <https://aip.scitation.org/doi/full/10.1063/1.4937166>.
However, I will reiterate David's initial comments that if you are mainly
interested in the hydration free energy of a bunch of small molecules then
you are probably better off using an implicit solvent model with a
correspondingly good quantum method. An implicit solvent model will allow
you to get a lot more bang-for-your-buck and you will avoid a lot of the
hiccups that arise in the MM->QM/MM transformation (needing to correct for
the change of the zero-in-energy, the fact that LJ terms are not changing
with the QM charge distribution for QM atoms, MM water has no polarization
response, the large errors that can be accrued over the different cycle
legs, etc). But if you are still interested in doing the TI or need to
have the explicit solvent then a point of clarification: Brian mentions
that you need to have MM charges that overlap with the QM charge
distributions for both the solvated system and in vacuum. That is true,
but it does not need to be the same set of charges. You can have different
charge sets for vacuum and solvated MM systems as well, which should help
the overlap in cases where the QM system is polarizing away from the
'standard' MM description.

Cheers,

On Thu, Nov 29, 2018 at 1:57 PM Brian Radak <brian.radak.gmail.com> wrote:

> The TI output is probably only directly usable for BAR if you do not use
> softcore potentials. Otherwise you will have to recompute energies.
>
> You are correct, the *magnitude* of the electrostatic component is almost
> always the dominant term, but the *variance* of the cavitation component is
> almost always the largest and thus contributes most to uncertainty (this is
> actually connected to the need for many lambda windows, since the TI
> derivative and variance are, within linear response, proportional). Suppose
> we have:
>
> dF = dF_elec + dF_vdW
> dF_elec = 100 +/- 0.5 kcal/mol
> dF_vDw = 10 +/- 2.0 kcal/mol
>
> It's true that there's an order of magnitude difference in the components,
> but the overall statistical error is still sqrt(0.5**2 + 2.0**2) = 2.1
> kcal/mol, which is dominated by the vdW term.
>
> The problem with the QM/MM charge distribution is that the polarization
> response is one-sided; the solvent never responds to the solute. This means
> the solute can over polarize, which might be way less accurate than the
> implicit (fixed) force field polarization. The other side of this is the LJ
> terms, which largely dictate the solute-solvent contact distance but also
> have no connection to the charge distribution. Arguably, the LJ terms have
> no connection to the QM method at all, since they are just grabbed from the
> nearest force field. You could maybe tune the parameters, but you pretty
> much need an experimental reference for this and it would be difficult to
> judge transferability.
>
> I'm not sure there are any popular codes that can handle QM/MM with
> polarization. I've seen this around for both CHARMM-style Drude and AMOEBA,
> but I think the results have been mixed. It would be interesting.
>
> Lastly, I think Adrian meant to imply that you need *two* endpoint
> MM->QM/MM corrections, since the difference in zero of energy needs to be
> cancelled out in both directions (otherwise you're just adding a large
> "random" number to your free energy). That means you need a gas phase
> correction also. That actually might be one of the bigger difficulties with
> the approach - you require overlap of the MM charges with both the solvated
> QM/MM and the vacuum QM. That would especially be a problem for charged
> molecules with large basis sets where the QM description actually gets
> problematic.
>
> BKR
>
>
>
>
> On Thu, Nov 29, 2018 at 2:31 PM Adrian Roitberg <roitberg.ufl.edu> wrote:
>
> > UGH!
> >
> > Indeed, Lee Woodcock, sorry !
> >
> > Adrian
> >
> > On 11/29/18 2:23 PM, Goetz, Andreas wrote:
> > > I believe Adrian meant papers by Lee Woodcock.
> > >
> > > I have not tested myself but Amber should be able to do what you
> intend.
> > >
> > > All the best,
> > > Andy
> > >
> > > —
> > > Dr. Andreas W. Goetz
> > > Assistant Research Scientist
> > > San Diego Supercomputer Center
> > > Tel: +1-858-822-4771
> > > Email: agoetz.sdsc.edu
> > > Web: www.awgoetz.de
> > >
> > >> On Nov 29, 2018, at 10:46 AM, Adrian Roitberg <roitberg.ufl.edu>
> wrote:
> > >>
> > >> One option is to create a thermo cycle where you compute the free
> energy
> > >> difference using a 'good' force field, and then correct the end point
> MM
> > >> --> QM/MM using TI changing the "Hamiltonian". The advantage is that
> you
> > >> do not need to make anything appear or disappear in QM. The
> disadvantage
> > >> is that you need the phase space sampled by MM and QM/MM to have a
> > >> reasonable overlap. See some recent paper by Lee Westwood for
> strategies
> > >> for this way of doing things.
> > >>
> > >> Adrian
> > >>
> > >> On 11/29/18 1:35 PM, Braden Kelly wrote:
> > >>> I do not see why I wouldnt use linear coupling for the electrostatic
> > part. I do this for regular classical MD free energies. The softcore LJ
> is
> > the same calculation for both MM and QM/MM is it not? So in theory, so
> long
> > as Amber can calculate for each lambda window the scaled coulomb
> > contribution to potential energy for the current window, and each
> > neighboring window(while still exploring phase space in the current given
> > lambda), along with the LJ contribution using a softcore potential, BAR
> and
> > TI should work. I guess my original question reflects this... Is amber
> set
> > up to do this? From the responses I have received, it appears to be
> capable
> > of doing TI and BAR provided I use a linear scaling for the coulomb and a
> > softcore for the LJ, which is also what I would do for a classical MD
> free
> > energy calculation.
> > >>>
> > >>>
> > >>>> From my experience nearly all of the Free Energy of inserting a
> > molecule into a solution, comes from the electrostatic contributions. LJ
> > requires more windows, but contributes much less. An issue with classical
> > FF is that aside from LJ parameters, their parameters come largely from
> QM
> > calculations done on a single molecule, in the gas phase (ideal gas). The
> > results are fit to spring constants and called a force-field. Generally
> > charges are fixed and forced to be symmetrical. I do not see how this
> can
> > be expected to be more accurate than doing QM on the molecule while it is
> > in the liquid phase while allowing the electronic contribution to be
> > polarized by the explicit and moving environment - unless you are using a
> > much lower quality of QM, or you spend some extra time parameter fitting
> or
> > refining the FF with a very high level of QM. I am not a FF developer
> and
> > everything I have said is my interpretation of what is done, please feel
> > free to criticize it! I prefer being corrected to being wrong...
> > >>>
> > >>> I am interested in a method that is as accurate as possible without
> > needing fitting to match experiment as is done with classical FF's since
> I
> > am interested in molecules which do not have experimental results to fit
> > to. This is why I am interested in QM/MM, I believe that especially if I
> > can include the first and hopefully second, hydration spheres in the QM
> > part of the calculation, I will get highly accurate results that are
> > subject to computational resources(level of theory) rather than FF
> > parameters. Perhaps I will need to also use a polarizable water model
> too.
> > >>>
> > >>>
> > >>> The point on water models is a very good one. First I would like to
> > get the most accurate results for given water models, and then compare
> > which model best mimics my reality. But I must do step 1 correctly,
> before
> > doing step 2.
> > >>>
> > >>>
> > >>> Braden
> > >>>
> > >>> ________________________________
> > >>> From: Brian Radak <brian.radak.gmail.com>
> > >>> Sent: Thursday, November 29, 2018 10:18:06 AM
> > >>> To: amber.ambermd.org
> > >>> Cc: david.case.rutgers.edu
> > >>> Subject: Re: [AMBER] QMMM free energy calculations for alchemical
> > changes using a coupling parameter in amber 18
> > >>>
> > >>> If you are doing TI with linear coupling then the data output is
> > identical
> > >>> to what would be needed for BAR/MBAR.
> > >>>
> > >>> Convergence for pure QM/MM transformations would almost certainly
> have
> > to
> > >>> be worse unless the conformations and/or charge distributions of the
> MM
> > >>> endpoints were substantially different. MM also has the distinct
> > advantage
> > >>> of being cleanly additive for staged decoupling.
> > >>>
> > >>> The other question is whether or not QM/MM solvation models are even
> > more
> > >>> accurate then MM solvation models. The former certainly were never
> > >>> calibrated to match experiment in any way. If anything, I would
> > *expect*
> > >>> them to be less quantitatively accurate. There have also been
> > questions for
> > >>> about 10-15 years now as to whether TIP3P or TIP4P models are "good"
> > >>> solvation models (they make good pure liquid models, but that's not
> at
> > all
> > >>> the same thing). Why would a more accurate (really, just "higher
> > level")
> > >>> electrostatic representation suddenly improve things? It was flawed
> in
> > some
> > >>> specific ways (see the erratum), but I would recommend reading David
> > >>> Mobley's comprehensive study of charge and solvent models for
> solvation
> > >>> free energies (one of the JPCs I think?).
> > >>>
> > >>> Cheers,
> > >>> BKR
> > >>>
> > >>>
> > >>>
> > >>> On Wed, Nov 28, 2018 at 4:15 PM Braden Kelly <bkelly08.uoguelph.ca>
> > wrote:
> > >>>
> > >>>> Thanks to both David and Brian for answering my question.
> > >>>>
> > >>>>
> > >>>> I will give TI a shot and be careful. I won't know if semi-empirical
> > will
> > >>>> work for my system until I try it. I have just had little to no luck
> > >>>> finding a software that would let me try it.
> > >>>>
> > >>>>
> > >>>> I used to be very excited about the end state correction QM/MM since
> > FEP
> > >>>> is always a little bit exciting, however am a little more guarded
> now.
> > >>>> Sampling and convergence is an issue, and there are cases where the
> > QM/MM
> > >>>> correction made the result deviate further from experiment than the
> > >>>> original MD free energy. From what I can tell, it is completely
> doable
> > >>>> using Amber, and conveniently, Amber looks like it also takes care
> of
> > the
> > >>>> QM reference state conversion. So with a bit of python or bash post
> > process
> > >>>> scripting, amber should work well for this option.
> > >>>>
> > >>>>
> > >>>> I look forward to when Amber and other software can do double sided
> > >>>> alchemical changes (BAR/MBAR) using QM/MM.
> > >>>>
> > >>>>
> > >>>> Explicit methods for calculating chemical potentials, which can then
> > be
> > >>>> used in a classical thermodynamic framework to predict nearly
> > anything of
> > >>>> chemical use, are my preferred choice, but I should explore all
> > options,
> > >>>> and will take a deeper look into implicit solvent models.
> > >>>>
> > >>>>
> > >>>> Thanks,
> > >>>>
> > >>>>
> > >>>> Braden
> > >>>>
> > >>>> ________________________________
> > >>>> From: David Case <david.case.rutgers.edu>
> > >>>> Sent: Tuesday, November 27, 2018 9:51:46 PM
> > >>>> To: AMBER Mailing List
> > >>>> Subject: Re: [AMBER] QMMM free energy calculations for alchemical
> > changes
> > >>>> using a coupling parameter in amber 18
> > >>>>
> > >>>> On Tue, Nov 27, 2018, Braden Kelly wrote:
> > >>>>> Amber documentation covers Free Energy calculations, such as
> > >>>>> thermodynamic integration and using other methods such as BAR. It
> > >>>>> independently discusses QM/MM as well. However free energy
> > calculations,
> > >>>>> specifically alchemical changes in which a molecule is decoupled
> > from the
> > >>>>> system, or mutated into another molecule, using a lambda parameter
> > that
> > >>>>> changes from 0 to 1 or vice versa, are not discussed in the context
> > of
> > >>>>> QM/MM.
> > >>>> In sander, one *can* put TI and QM/MM together; special care would
> > need to
> > >>>> be taken with softcore LJ terms (and use of softcore electrostatics
> > would
> > >>>> not work at all.) I don't see any way to use BAR (but I might be
> > missing
> > >>>> something).
> > >>>>
> > >>>> The downside is that one needs to be able to afford the QM/MM
> > >>>> calculation in the first place, and to carry out significant
> > simulations
> > >>>> at a variety of lambda points. This requires a *lot* of
> computation,
> > >>>> unless you happen to have a system where semiempirical or DFTB
> > >>>> Hamiltonians are accurate enough for your purposes.
> > >>>>
> > >>>> It is for this reason that a big portion of current research in this
> > >>>> area explores the suggestion made by Brian: carry out a
> transformation
> > >>>> at the MM level, then apply MM->QM/MM "corrections" at the end
> points.
> > >>>> In favorable circumstances, this can involve many fewer QM
> > calculations,
> > >>>> which generally are by far the most expensive parts of the
> simulation.
> > >>>>
> > >>>>> I am specifically interested in hydration free energies of small
> (<50
> > >>>>> atoms) organic molecules.
> > >>>> If you want just the hydration free energies, you might be better
> off
> > >>>> which implicit solvent models, such as the various flavors of
> > >>>> "Minnesota" models (such as SMx and its many variants.) Using
> > explicit
> > >>>> solvent models (and a program like Amber) would mainly be of
> interest
> > if
> > >>>> you want the difference in solvation energy between two
> environments,
> > >>>> say pure water, and a protein/water environment.
> > >>>>
> > >>>> ....good luck....dac
> > >>>>
> > >>>>
> > >>>> _______________________________________________
> > >>>> AMBER mailing list
> > >>>> AMBER.ambermd.org
> > >>>>
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwICAg&c=pZJPUDQ3SB9JplYbifm4nt2lEVG5pWx2KikqINpWlZM&r=JAg-KQEjdZeg_E8PHDDoaw&m=FTPeqdqETo2lr7MW_xvsTnmEgSCvsh7L4HyaImIRAgE&s=yfuvwfVuiLnMLvwJd5BNqf76japdXEJOSMf1mf30yA0&e=
> > >>>> _______________________________________________
> > >>>> AMBER mailing list
> > >>>> AMBER.ambermd.org
> > >>>>
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwICAg&c=pZJPUDQ3SB9JplYbifm4nt2lEVG5pWx2KikqINpWlZM&r=JAg-KQEjdZeg_E8PHDDoaw&m=FTPeqdqETo2lr7MW_xvsTnmEgSCvsh7L4HyaImIRAgE&s=yfuvwfVuiLnMLvwJd5BNqf76japdXEJOSMf1mf30yA0&e=
> > >>>>
> > >>> _______________________________________________
> > >>> AMBER mailing list
> > >>> AMBER.ambermd.org
> > >>>
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwICAg&c=pZJPUDQ3SB9JplYbifm4nt2lEVG5pWx2KikqINpWlZM&r=JAg-KQEjdZeg_E8PHDDoaw&m=FTPeqdqETo2lr7MW_xvsTnmEgSCvsh7L4HyaImIRAgE&s=yfuvwfVuiLnMLvwJd5BNqf76japdXEJOSMf1mf30yA0&e=
> > >>> _______________________________________________
> > >>> AMBER mailing list
> > >>> AMBER.ambermd.org
> > >>>
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwICAg&c=pZJPUDQ3SB9JplYbifm4nt2lEVG5pWx2KikqINpWlZM&r=JAg-KQEjdZeg_E8PHDDoaw&m=FTPeqdqETo2lr7MW_xvsTnmEgSCvsh7L4HyaImIRAgE&s=yfuvwfVuiLnMLvwJd5BNqf76japdXEJOSMf1mf30yA0&e=
> > >> --
> > >> Dr. Adrian E. Roitberg
> > >> University of Florida Research Foundation Professor
> > >> Department of Chemistry
> > >> University of Florida
> > >> roitberg.ufl.edu
> > >> 352-392-6972
> > >>
> > >>
> > >>
> > >> _______________________________________________
> > >> AMBER mailing list
> > >> AMBER.ambermd.org
> > >>
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwIGaQ&c=pZJPUDQ3SB9JplYbifm4nt2lEVG5pWx2KikqINpWlZM&r=JAg-KQEjdZeg_E8PHDDoaw&m=hXCfXj8ZVOd_i_SgzKPVZaN9SDtEyKs9dTBmLJYB4a0&s=Y8-cxYIYtX2Wi4-ixjdVgwmd3EY2bkDe8SA5S8UCa70&e=
> > > _______________________________________________
> > > AMBER mailing list
> > > AMBER.ambermd.org
> > >
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwIGaQ&c=pZJPUDQ3SB9JplYbifm4nt2lEVG5pWx2KikqINpWlZM&r=JAg-KQEjdZeg_E8PHDDoaw&m=hXCfXj8ZVOd_i_SgzKPVZaN9SDtEyKs9dTBmLJYB4a0&s=Y8-cxYIYtX2Wi4-ixjdVgwmd3EY2bkDe8SA5S8UCa70&e=
> >
> > --
> > Dr. Adrian E. Roitberg
> > University of Florida Research Foundation Professor
> > Department of Chemistry
> > University of Florida
> > roitberg.ufl.edu
> > 352-392-6972
> >
> >
> >
> > _______________________________________________
> > AMBER mailing list
> > AMBER.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber
> >
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>


-- 
Erich Kuechler
erich.kuechler.gmail.com
erich.kuechler.msl.ubc.ca
====================================================
Mayor and Gsponer Groups - Postdoctoral Research Fellow in Computational
Biology
Michael Smith Laboratories - The Centre for High-Throughput Biology
The University of British Columbia
Doctor of Philosophy - Chemistry, Prof Darrin York
University of Minnesota - Twin Cities, Department of Chemistry
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Nov 29 2018 - 17:30:02 PST
Custom Search