[gmx-users] REMD and GBSA

[Mark Abraham]

On 18/10/2011 4:58 AM, Ben Reynwar wrote:
> On Fri, Oct 14, 2011 at 9:11 AM, Mark Abraham<Mark.Abraham at anu.edu.au>  wrote:
>> On 14/10/2011 10:12 AM, Ben Reynwar wrote:
>>> Hi gromacs list,
>>>
>>> I'm about to start some REMD simulations using generalized Born
>>> solvent on a protein of about 5000 atoms.  I have two questions, the
>>> first of which is about gromacs, the second more about REMD in
>>> general.
>>>
>>> (1)
>>> I'm getting some pretty ugly energy drift (300K->500K in 1 ns) for an
>>> NVE MD test simulation using a 2 fs time step.  It goes away if I use
>>> 1 fs, however I was under the impression that 2 fs is normally OK.  I
>>> was wondering whether that could be caused by the use of the cut-off
>>> method which is required with the coloumb and VdW interactions when
>>> using GBSA?  Or perhaps I'm doing something else wrong.  I'll include
>>> the mdp file I'm using at the bottom of the email in case anyone feels
>>> like pointing out my foolishness to me.
>> Drift is normal if you use 2fs with hbond constraints. all-bond constraints
>> are necessary for 2fs.
>>
> Great. Thank you.
>
>>> (2)
>>> I've been analyzing some data from an REMD simulation by my
>>> predecessor and see very slow replica flow rates.  They are about two
>>> orders of magnitude smaller than the idealized rate of the exchange
>>> attempt frequency multiplied by the acceptance fraction (exchanges are
>>> attempted every 2 ps with a 0.4 acceptance fraction).  If I look at
>>> the energy distribution for a given replica/temperature combination
>>> over a time scale of around 1 ns, it is clearly shifted from the
>>> average energy distribution for that temperature.  The timescale for
>>> changes in this energy shift is around 10 ns.  My current theory for
>>> the slow rate of replica flow is that the slow fluctuations in the
>>> energy of the protein are limiting replica flow, since a replica with
>>> lower than average energy will tend to remain at the bottom of the
>>> temperature range, while those with higher than average energies will
>>> tend to remain at the top.  Has anyone else observed this kind of
>>> behavior?  Is my reasoning wrong in any obvious way?
>> A replica with "lower than average energy" *for that temperature* will tend
>> to drift down the temperature ladder in favour of another.
>>
>> One can observe blockages in replica flow. If all the replicas below a given
>> temperature are in regions of configuration space that cannot access PE high
>> enough to have a significant chance of exchanging above that temperature,
>> then flow does not occur (and vice-versa, of course). If one were to sample
>> a FES that had two minima that should be sampled in a 2:1 ratio, but started
>> from a 1:1 ratio and did not have a high enough temperature range to cross
>> the barrier, then the exchange acceptance rate can look good when nothing
>> useful is occurring - the observation will necessarily be that these minima
>> are equally likely. The two groups are actually engaging in disjoint flow,
>> and one needs to look at metrics other than the acceptance rate to observe
>> it. The only way to deal with such a bottleneck is to have replicas at a
>> high enough temperature that both groups can exchange to those temperatures
>> - only now can barrier crossing occur. These kinds of phenomena can
>> certainly occur over short time scales in localized regions of configuration
>> and temperature space.
>>
>> Mark
>>
> Do you know if anyone has done any studies looking at replica flow in
> well-defined, comparatively low-dimensional landscapes to get a
> qualitative feel for these kinds of effects?  It would be interesting
> to see what effect replica exchange settings can have on replica flow
> beyond the simple random walk models, when you take into account the
> fact that different regions of configuration space could have
> different potential energy distributions for the same temperature.
>

See for example Walter Nadler, Jan H. Meinke, and Ulrich H. E. Hansmann 
"Folding proteins by first-passage-times-optimized replica exchange", 
PHYSICAL REVIEW E 78, 061905, 2008 (and related work)

Mark