[gmx-users] optimum acceptance ratio for REMD

[David van der Spoel]

On 2012-01-12 19:42, Ben Reynwar wrote:
> Hi all,
>
> I am currently puzzled by the emphasis on using an acceptance ratio of
> around 0.2 for replica exchange simulations in the literature.
>
This is bogus indeed.

You are correct. Many people prefer not to think.

What matters is that the time between exchanges is larger than the 
relaxation, which hence is system dependent, and that you have 
sufficient number of exchanges to equilibrate the T-dependent properties 
(which can be challenging).

Our old paper is one of the few where this succeeded for a peptide 
because it is so small :).

Marvin Seibert, Alexandra Patriksson, Berk Hess and David van der Spoel: 
Reproducible polypeptide folding and structure prediction using 
molecular dynamics simulations J. Mol. Biol. 354 pp. 173-183 (2005)

David van der Spoel and M. Marvin Seibert: Protein Folding Kinetics and 
Thermodynamics from Atomistic Simulations Phys. Rev. Lett. 96 pp. 238102 
(2006)
> This optimum ratio is, as far as I understand, derived assuming that
> the exchange attempt frequency is constant and that the protein
> dynamics are not limiting the diffusion of a replica up and down
> the temperature ladder.
>
> In a real system there will be some time-scale associated with the
> relaxation of the protein as it moves from one temperature to another.
> The rate of diffusion of the replica on the temperature ladder could
> potentially be limited by these protein dynamics or by the exchange
> frequency.
>
> If we are limited by the exchange frequency, then it would make sense
> to decrease the interval between exchange attempts to the same
> time-scale as the potential energy correlation time, and to choose the
> number of replicas such that the acceptance ratio was around 0.2.
>
> If we are limited by the protein dynamics then the exchange frequency
> doesn't matter, so the acceptance ratio should have no effect (as
> long, of course, as it isn't so low that the exchange frequency
> becomes limiting).  This would mean for simulations of large proteins where
> the protein dynamics are limiting, we can slash the number of replicas
> required for an efficient REMD simulation.
>
> However, I haven't seen any simulations in the literature using very small
> acceptance ratios, which makes me suspicious of my logic.  Does anyone have
> any thoughts on this?
>
> Cheers,
> Ben


-- 
David van der Spoel, Ph.D., Professor of Biology
Dept. of Cell & Molec. Biol., Uppsala University.
Box 596, 75124 Uppsala, Sweden. Phone:	+46184714205.
spoel at xray.bmc.uu.se    http://folding.bmc.uu.se