Re: DOF during alchemical simulations

From: Grace Brannigan (gracebrannigan_at_gmail.com)
Date: Mon Nov 30 2015 - 14:45:55 CST

Hi Brian,

On Mon, Nov 30, 2015 at 1:30 PM, Brian Radak <bradak_at_anl.gov> wrote:

>
> Sorry, yes, this is what I meant to convey. The code currently seems to
> favor the "alchDecouple" on scenario.
>

Yes, life is just easier if you're doing decoupling rather than
annihilation - for several other reasons similar to the one you brought up
here...

> I only meant that a periodic molecule or set of atoms is not an ideal gas
> because there is a configurational component. In that case though, there
> *is* a coupling between the two systems since they share a box and both
> push on.
>

Well, technically an ideal gas also has a configurational component, it's
just unweighted - but I think I know what you mean. It probably goes
without saying that if small molecules or highly localized groups are
interacting with their images then the simulation has bigger problems - so
in most cases interactions among periodic images won't contribute any
theoretical or actual differences between using alchDecouple On & Off.

> I agree that the error is likely negligible. This is all an exercise in
> fastidiousness and I doubt any changes would affect anyone's results in a
> hugely discernible way.
>

Great description of most issues that come up with FEP :-) The best case
scenario is realizing the correction is negligible sooner rather than
later... In this case the only straightforward fix is probably setting
alchDecouple On.

> I'm not entirely repulsed by the sudden introduction of degrees of freedom
> when moving away from lambda = 0,1, as this is more in line with single
> coordinate transformations, just with a alchemical path. Although this
> might also add interesting shifts to the TI derivative.
>
>

> HTH,
> Grace
>
> On Mon, Nov 30, 2015 at 10:36 AM, Brian Radak <bradak_at_anl.gov> wrote:
>
>> I'm glad this is intriguing rather than a non-issue. I agree that
>> whatever systematic errors are present are likely quite negligible.
>> Nonetheless, they can still be discerned from simple tests and I dislike
>> being incorrect when the right answer can be easily achieved.
>>
>> My thought was that the degrees of freedom ought to depend on the
>> decoupling scheme (alchdecouple on/of), as this determines whether or not
>> the annihilated atoms see their images (exist as a periodic "gas") or not
>> (are an ideal gas molecule). Does it make sense for ideal gas degrees of
>> freedom to impact the pressure? My first thought would be that they should
>> not impact sampling at all; I believe manipulations with ideal gas
>> partition functions ought to confirm this is true.
>>
>> Would it make sense to have different behavior when alchDecouple is on or
>> off? This would only meaningfully differ at alchLambda = 0,1, as the
>> intervening values are totally arbitrary, so long as sampling is not
>> grossly impacted.
>>
>> Brian
>>
>>
>> On 11/27/2015 03:31 AM, Jérôme Hénin wrote:
>>
>> On 26 November 2015 at 19:28, Aron Broom < <broomsday_at_gmail.com>
>> broomsday_at_gmail.com> wrote:
>>
>>> This is really interesting. My knowledge of alchemical transformations
>>> in limited, but given their successes I'd like to understand more (and will
>>> happily be corrected on my errors in thinking).
>>>
>>> If you leave those degrees of freedom in, then the end-point simulations
>>> are actually different than a similar simulation of that system where you
>>> aren't doing an alchemical transformation. That raises for me a kind of
>>> intuitive red-flag, which I think is the same point you are making?
>>>
>>
>> I agree with you on this. This is among the terms that we neglect when
>> doing alchemical calculations in an isobaric simulations. If you decouple n
>> particles among N and have a barostat set at pressure P0, you will generate
>> an ensemble for the (N-n) particles at pressure P = P0 - Pn, where Pn is
>> the kinetic pressure from just n particles at the given volume and
>> temperature. If I get my orders of magnitude right: decoupling one particle
>> in a thousand from a condensed phase will underestimate the pressure by (on
>> the order of) 1 bar. That's something I can live with: I can say worse
>> things about my simulations.
>>
>>
>>> But on the other hand, if at the end-points you suddenly eliminate those
>>> degrees of freedom completely, doesn't that create a discontinuity in the
>>> transformation, which is a bad thing and source of much misery?
>>>
>>
>> Unless you do TI, it's not a problem in and of itself: other estimators
>> explicitly give FE differences between discrete states. The tricky part may
>> be to account for that explicitly in the free energy estimator.
>>
>> Probably an idiotic question from someone with limited physics
>>> understanding, but I suppose non-integer degrees of freedom are disallowed
>>> (assuming similar fractional counting of mass and velocity)?
>>>
>>
>> Nothing prevents us from using a fractional number when calculating
>> kinetic pressure, although it doesn't have much physical meaning. That's
>> pretty much the spirit of alchemical transformations. Again, I'd be totally
>> happy with it if the estimators were rewritten with that in mind.
>>
>> Jerome
>>
>>
>>
>>> On Thu, Nov 26, 2015 at 1:00 PM, Jérôme Hénin < <jerome.henin_at_ibpc.fr>
>>> jerome.henin_at_ibpc.fr> wrote:
>>>
>>>> Brian,
>>>>
>>>> I might be missing something, but I'd say the degrees of freedom of
>>>> non-interacting particles should be counted for the purpose of kinetic
>>>> pressure calculation.
>>>>
>>>> Jerome
>>>>
>>>> On 25 November 2015 at 17:31, Brian Radak < <bradak_at_anl.gov>
>>>> bradak_at_anl.gov> wrote:
>>>>
>>>>> After some griping about this, I've finally implemented a
>>>>> (preliminary) correction to the Lennard-Jones tail correction that accounts
>>>>> for alchemical modifications. Once this is integrated with other
>>>>> improvements to the alchemical code, I hope this will become part of the
>>>>> 2.11 release.
>>>>>
>>>>> However, I recently noticed that a similar problem crops up in the
>>>>> degrees of freedom calculation. That is, alchemical atoms get counted at
>>>>> the endpoints even when they are only ideal gas particles. This was obvious
>>>>> when I started double checking single coordinate endpoint energies and
>>>>> pressures with dual coordinate alchemical energies and pressures; that is,
>>>>> the energies match but the pressures do not quite match.
>>>>>
>>>>> The error is admittedly much less than 0.1%, as multiplying a "more
>>>>> different" large number by a small number is still just another "kind of
>>>>> large" number. Nonetheless, one could view this as an error in the
>>>>> specified target pressure for an alchemical simulation (i.e. the pressure
>>>>> you input is not the pressure you simulate). Then again, this behavior
>>>>> might be exactly what one is expecting, depending on how one draws the
>>>>> thermodynamic cycle.
>>>>>
>>>>> I guess my question for the community is, does this matter? How do
>>>>> people expect degrees of freedom to be determined? Do people usually draw
>>>>> their cycles such that non-interacting particles should not contribute?
>>>>> This might not be the case, for example, in ligand binding calculations
>>>>> where the ligand continues to interact with its own images (although in
>>>>> that case, one essentially has two simulations going at the same time when
>>>>> the ligand is decoupled).
>>>>>
>>>>> Brian
>>>>>
>>>>>
>>>>> --
>>>>> Brian Radak
>>>>> Theta Early Science Program Postdoctoral Appointee
>>>>> Leadership Computing Facility
>>>>> Argonne National Laboratory
>>>>>
>>>>> 9700 South Cass Avenue
>>>>> Building 240, 1.D.16
>>>>> Lemont, IL 60439-4871
>>>>> Tel: 630/252-8643
>>>>> email: <bradak_at_anl.gov>bradak_at_anl.gov
>>>>>
>>>>>
>>>>
>>>
>>>
>>> --
>>> Aron Broom M.Sc
>>> PhD Student
>>> Department of Chemistry
>>> University of Waterloo
>>>
>>
>>
>> --
>> Brian Radak
>> Theta Early Science Program Postdoctoral Appointee
>> Leadership Computing Facility
>> Argonne National Laboratory
>>
>> 9700 South Cass Avenue
>> Building 240, 1.D.16
>> Lemont, IL 60439-4871
>> Tel: 630/252-8643
>> email: bradak_at_anl.gov
>>
>>
>
>
> --
> Grace Brannigan, Ph.D.
> Assistant Professor
> Center for Computational and Integrative Biology (CCIB) &
> Department of Physics
> Rutgers University, Camden, NJ
> (856)225-6780
> http://branniganlab.wordpress.com
>
> --
> Brian Radak
> Theta Early Science Program Postdoctoral Appointee
> Leadership Computing Facility
> Argonne National Laboratory
>
> 9700 South Cass Avenue
> Building 240, 1.D.16
> Lemont, IL 60439-4871
> Tel: 630/252-8643
> email: bradak_at_anl.gov
>
>

This archive was generated by hypermail 2.1.6 : Tue Dec 27 2016 - 23:21:35 CST