From: Jason Swails <>
Date: Wed, 05 Nov 2014 12:05:07 -0500

On Wed, 2014-11-05 at 16:20 +0000, Mele N. wrote:
> Dear Amber user,
> I am currently using Temperature REMD simulation and I would like to
> use H-REMD to improve my knowledge.
> However, I can't understand how to perform H-REMD in AMBER. With
> T-REMD we specify as an input file a set of temperatures (generated by
> gromacs generator in my case) so I would like to understand how do we
> choose the number of replicas?

Guesstimate. Run a short simulation and take a look at the exchange
acceptance rates and adjust your replica number (and/or spacing) based
on the results.

Note that T-REMD is rather unique in this regard. The replica spacing
comes from potential energy distributions, whose fluctuations (i.e., the
width of the potential energy distributions) are based in statistical
mechanics and vary inversely with the square root of the number of
particles. I think there may be _some_ empiricism in the formula the
GROMACS temperature generator uses, but at the end of the day you can
get a pretty good prediction of the exchange probabilities based on just
the number of degrees of freedom in your system.

There is no analog to this in H-REMD, where acceptance rates in state
space depend on the complex nature of the underlying Hamiltonian and
exchange dimension. What's more, the shape of the underlying free
energy surface in state space is not the same (or even similar) for
every system, so the "optimum" replica number and spacing you choose for
one system is not necessarily transferable to another. (Note that
"optimum" here can have different exact definitions, but that does not
affect the validity of my statement).

Consider an umbrella replica exchange simulation -- the umbrella
coordinate distributions will be far narrower in regions near the
transition state where the free energy surface is the steepest, so you
will need more replicas in that region to avoid introducing a bottleneck
in state space sampling.

> I tried to find paper or tutorial on it but I couldn't managed, if you
> have any advices it would be appreciated.

Have a look at While
this paper doesn't answer your exact question and is specific to
umbrella sampling, it should help yield some of the insight needed to
figure it out yourself. It should also be possible, if not
straightforward, to extend these ideas to other forms of replica
exchange (like alchemical replica exchange free energy perturbation,
pH-replica exchange, etc.)

Another way of thinking about this is that you need _some_ knowledge of
the underlying surface in state space in order to pick the "best"
locations for your replicas (as well as how many replicas will be
needed). With T-REMD, you have knowledge a priori of what this surface
looks like. With H-REMD, in the most general case, you do not. The
best way to estimate the shape of that surface is to start with your
best guess of where you should put replicas and just give it a try.


Jason M. Swails
Rutgers University
Postdoctoral Researcher
AMBER mailing list
Received on Wed Nov 05 2014 - 09:30:03 PST
Custom Search