Re: [AMBER] REMD simulation

From: Niel Henriksen <niel.henriksen.utah.edu>
Date: Fri, 15 Mar 2013 01:39:13 +0000

Sudipta,

We have run quite a few REMD simulations in explicit solvent.
They are computationally expensive, but we do it because our
RNA simulations perform badly with implicit solvent.

We just published a paper which discusses some of the hurdles
with this approach, specifically convergence. If you want to do
explicit solvent REMD, take a look:

http://pubs.acs.org/doi/abs/10.1021/jp400530e

Good luck,
--Niel
________________________________________
From: sudipta [sudipta.mml.gmail.com]
Sent: Thursday, March 14, 2013 5:51 PM
To: AMBER Mailing List
Subject: Re: [AMBER] REMD simulation

Hi Jason,

Thank you very much for the answer in detail. I truly appreciate your
comments.

Best
Sudipta

On Thu, Mar 14, 2013 at 7:43 PM, Jason Swails <jason.swails.gmail.com>wrote:

> On Thu, Mar 14, 2013 at 7:07 PM, sudipta <sudipta.mml.gmail.com> wrote:
>
> > Hi Niel,
> >
> > Thank you very much for your reply. Yeah, I noticed that most of the
> people
> > had used NVT instead of NPT. Nobody has not put any straightforward
> reason
> > for choosing NVT over NPT.
>
>
> Convenience and simplicity. The exchange probability is currently
> calculated from basically a ratio of Boltzmann probabilities between the
> two states in adjacent replicas. Boltzmann probabilities are valid for
> NVT, only. To support NPT with REMD, a PV term needs to be added, I think.
>
>
> > However, is the simulation of protein system at
> > high temperature in NVT ensemble reasonable? When we increase the
> > temperature of such system then the protein molecule will be expanded and
> > it looks for some space. However, NVT ensemble doesn't compromise any
> > space. I am curious about this.
> >
>
> When fluctuations are not important, all ensembles are effectively
> identical. It is only when these fluctuations begin to become large (at,
> for instance, phase transitions), that you need to be careful about your
> ensemble. However, the purpose of the high-temperature replicas is NOT to
> see how the system behaves at that temperature. In fact, many T-REMD
> simulations have replicas at temperatures over the boiling point of water
> at 1 bar.
>
> To properly model this behavior, you would naturally need to use NPT... but
> you don't want to model the behavior of boiling proteins (if you did, then
> a force field parametrized at 300K is _not_ the way to go). The only
> purpose of using high-temperature replicas is to generate a series of
> structures that 'might' be relevant at a lower temperature that are hard to
> reach there due to large barriers in phase space.
>
> The fact that high-temperature replicas are nonsensical due to the force
> field not being validated at such temperatures as well as the issues you
> raise due to our choice of ensemble are actually irrelevant. Because the
> exchange probability obeys detailed balance, you are assured that your
> low-temperature replicas are thermodynamically correct. The
> high-temperature replicas are used only to enhance sampling and serve
> little use after the simulation is over (you can use them for some stuff,
> like reweighting, but are often just discarded).
>
> In fact, if you ran NPT, then you would actually _hurt_ your exchange
> success rate if you came anywhere close to the boiling point of water,
> since the volume would expand too much. So IMO, NVT is the better choice
> for most REMD simulations.
>
> Another issue is use of explicit solvent. Most of REMD study has been
> > performed in implicit solvent environment. Is there any inherent reason
> > behind this except computational cost.
> >
>
> Computational cost is the only reason. But the computational cost
> increases for two reasons. First, more particles means slower simulations,
> and the viscosity of water molecules hinders conformational sampling in
> explicit solvent compared to implicit solvent -- so you need longer
> simulations that run slower.
>
> Even worse with REMD, though, is that the width of your potential energy
> distribution scales as 1/sqrt(N) where N is the # of particles. With
> explicit solvent (or any large system), the number of particles becomes
> large, so the potential energy distributions become narrow. Since these
> distributions must overlap in order for REMD to be effective, you need to
> place replicas much closer together. Therefore, in order to span the same
> temperature range in explicit solvent, you often need many, many more
> replicas.
>
> That is why T-REMD simulations are typically (although not always) run with
> implicit solvent unless the system size is rather small to begin with.
> (e.g., alanine dipeptide)
>
> HTH,
> Jason
>
> --
> Jason M. Swails
> Quantum Theory Project,
> University of Florida
> Ph.D. Candidate
> 352-392-4032
> _______________________________________________
> 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

_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Mar 14 2013 - 19:00:03 PDT
Custom Search