Re: [AMBER] Explicit vs Implicit solvent for REMD

From: Jason Swails <>
Date: Thu, 25 Jul 2013 12:17:36 -0400

On Thu, Jul 25, 2013 at 9:03 AM, Parker de Waal <>wrote:

> Hi Amber Users,
> I'm currently looking into using REMD to simulate the the effects of a
> single aa substitution and was wondering if anyone had any experience with
> implicit vs explicit solvent models. To my knowledge explicit provides a
> higher level of accuracy at the expense of increased computational demand,
> however I am unsure how the choice of explicit vs implicit would affect the
> results of REMD.

First off, it's important to realize that REMD, when implemented and used
properly, samples from exactly the same stationary distributions
(ensembles) as uncoupled, regular MD simulations run under the exact same
conditions. The only difference being that replicas can traverse state
space as well as its standard phase space in order to dodge and/or scale
high barriers. As a result, the difference between implicit and explicit
solvent in REMD, as far as sampling and accuracy, is exactly the same as it
is in standard MD. A simulation run out infinitely long without REMD will
give exactly the same results as an infinitely long

REMD simulation, the only difference being that the REMD simulations
should converge faster.

The importance of implicit vs. explicit solvent comes from the impact that
explicit water molecules has on the required number of replicas. For
T-REMD, specifically, exchange probabilities are calculated via:

Prob = exp(-(Beta_i - Beta_j)*(E_i - E_j))

Therefore, the exchange probability is only high if the temperature and/or
energy differences are small. The replicas must be spaced such that the
probability is high enough to give proper mixing (~20% acceptance rate is a
good target). A good way to see if you will get good exchange rates is to
look at the overlaps of the energy distributions between adjacent replicas.
 If there is 'good' overlap, then there will be a good chance that exchange
attempts will occur when the energy difference is small (or, even better,
negative -- those exchanges are always accepted).

The complication here is that the widths of the energy distributions scales
as 1/sqrt(N), where N is the number of particles in the system (standard
stat mech -- look for 'fluctuations' in your favorite text). Thus, as you
increase the number of particles, the widths of the potential energy
distributions decreases. This in turn requires that you use more replicas
spaced closer together in temperature-space to ensure good exchange rates.
 Since explicit solvent simulations tend to have a huge number of particles
compared to implicit solvent simulations, explicit solvent simulations
often require many more replicas for T-REMD than do implicit solvent
calculations to cover the same temperature range. You can play around with
this link to see the effect of explicit solvent on the number of necessary

Note, this is specific to T-REMD. Other versions of REMD (e.g., pH-REMD
and umbrella REMD) depend on fluctuations of different variables (like
total number of titrating protons for pH-REMD and the strength of the
harmonic restraints for umbrella REMD), so the same arguments do not
universally apply to "replica exchange" in general.


Jason M. Swails
Rutgers University
Postdoctoral Researcher
AMBER mailing list
Received on Thu Jul 25 2013 - 09:30:03 PDT
Custom Search