Re: Advisable limit on the number of quadrants a 2D ABF simulation can be split.

From: Jrme Hnin (jhenin_at_ifr88.cnrs-mrs.fr)
Date: Mon Mar 28 2011 - 09:42:33 CDT

Dear Ajasja,

[note: slightly revised answers]

On 28 March 2011 16:06, Ajasja Ljubetič <ajasja.ljubetic_at_gmail.com> wrote:
>
> Dear Jerome,
>
> Thank you for the comforting thought about umbrella sampling:)
>
> My system is an alpha helix made of 35 alanines, spin labelled in the middle
> (position 18) and put in a DMPC membrane. The spin label is a bulky
> side-chain, comparable to trypthopan in size.  The reaction coordinates are
> the theta and phi angles of the vector from the labels Ca atom to the last O
> atom.
>
> At first I wanted to spin label different positions on the peptide, but
> right now I don't think this is the best idea. If I run several 2D free
> energy scans at different positions, I will not get the offset between them
> (as each scan will have the zero set to it's global minimum). But If I
> define the membrane depth as another colvar and run what is effectively a 3D
> scan then the offsets between various 2D scans will be correct. The depth
> colvar would be the distance between the COM of the membrane and the C-alpha
> of the spin label.
> A simple 1D depth scan will probably not help either, because the other
> degrees of freedom (the theta and phi colvars) are to slow. Is my reasoning
> sound?

Indeed, if they are slow and tightly coupled to depth, they could
certainly be a problem.

> I think I have just enough computer power to perhaps converge a 3D
> ABF free energy scan.

It could be worth a try. If you do that, you'll have as much
experience with 3D ABF as I.

> two further questions:
>
> What did you mean by monitoring relaxation? (currently I just watch each ns
> how the PMF converges).

That is, indirectly looking at relaxation. I tend to look at sampling
histograms and colvar trajectories more than the PMF itself. Then, if
something looks odd, VMD is your friend.

> If I set a distanceZ colvar between (0, 0,0) and the COM of the membrane
> (using the phosphate heads), will this be handled correctly if half of the
> membrane passes into the "next" periodic cell? I think I should
> use forceNoPBC but I'm not sure how the coordinates are stored internally.

For the purpose of COM calculation, the atom group is assumed to be
"in one piece". ForceNoPBC will not change this. Once the restraint is
in place, however, it is unlikely that lipids drift far enough to be a
problem. You can always disable wrapAll to avoid nasty surprises when
restarting jobs.

Best,
Jerome

>
> On Fri, Mar 25, 2011 at 12:25, Jérôme Hénin <jhenin_at_ifr88.cnrs-mrs.fr>
> wrote:
>
>> Hi Ajasja,
>>
>> It's hard to have a gut feeling when you didn't tell us anything about
>>
>> your system. As a comparison though, remember that people who do
>>
>> umbrella sampling would probably chop up this space into at least a
>>
>> hundred windows, and suffer any convergence problems that would arise.
>>
>> The difference is, they would have no other choice (besides
>>
>> relinquishing umbrella sampling entirely). So whatever you end up
>>
>> doing, you can find comfort in the fact that it's better than umbrella
>>
>> sampling.
>>
>> Anyway, the questions are: are there degrees of freedom other than
>>
>> your RCs that can get stuck? (hint: the answer is yes) and then: how
>>
>> much do the RCs have to move about to give those a chance to get
>>
>> unstuck (and how long will it take)? That should give you an idea of a
>>
>> minimum size for the quadrants, and a minimum sampling time for each
>>
>> of them (of course that can and should be adjusted by monitoring
>>
>> relaxation in the simulations).
>>
>> Cheers,
>>
>> Jerome
>>
>> On 25 March 2011 10:52, Ajasja Ljubetič <ajasja.ljubetic_at_gmail.com> wrote:
>>
>> > Dear all,
>>
>> > Is there any advisable limit on how many quadrants a 2D torsional free
>>
>> > energy scan using ABF can be divided? Scaling the simulation by
>> > splitting
>>
>> > the ABF calculation has the advantage that each part runs independently,
>> > so
>>
>> > there are no scaling issues.
>>
>> > On the other hand one is wary of introducing artifacts by the artificial
>>
>> > wall constraints (as in figure 6b in [1]).
>>
>> > I'm asking because over the weekend I will have access to 512 cores. My
>>
>> > simulation has 16.000 atoms, so just throwing more cores at it will not
>> > be
>>
>> > very effective. (I will of course run the necessary scaling benchmarks
>>
>> > before I start)
>>
>> > Currently I'm spliting the ABF calculation into 8 parts. Would 36 parts
>> > be
>>
>> > too much? This would then be pieces of 6x6 points.(The bin width is 10
>>
>> > degrees and the range form -180 to 180).
>>
>> > I know this strongly depends on the system, but I'm asking for the "gut"
>>
>> > feeling:)
>>
>> > Thank you & best regards,
>>
>> > Ajasja Ljubetič,
>>
>> > Young reasercher,
>>
>> > Laboratory of biophysics,
>>
>> > Institute Jožef Štefan,
>>
>> > Ljubljana, Slovenia
>>
>> >
>>
>> > Christophe Chipot and Jérome Hénin, “Exploring the free-energy
>> > landscape
>>
>> > of a short peptide using an average force,” The Journal of Chemical
>> > Physics
>> > 123, no. 24 (2005): 244906.
>
>

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:56:50 CST