Re: Replica Exchange

From: Jason Swails (
Date: Tue Jan 19 2016 - 15:06:24 CST

On Tue, Jan 19, 2016 at 11:35 AM, Dr. Eddie <> wrote:

> HI all,
> This isn't a namd-only question but since RE is used by many namd users
> I'm hopeing someone can point me to a good reference. I have a simple
> question:
> Why is replica exchange more efficient than just heating a single system
> up and then cooling it when it gets stuck (one could use the same criteria
> for doing this as RE)?
> More precisely, why is it more efficient to have N-replicas at different
> temperatures rather than N-individual runs with different initial
> conditions that are heated/cooled as needed to get over potential barriers?

‚ÄčI think the alternative method you're referring to is expanded ensemble
techniques -- see

I think the main thing that T-REMD has going for it over expanded ensemble
techniques basically comes down to simplicity. You don't have to tune as
many parameters (that may be system-specific) in order to strike a balance
between making sure most of your sampling is at your T of interest while
still sampling ENOUGH at higher temperatures to jump barriers (note that
your simulation loses its reversibility if you only jump temperatures when
you detect you may have gotten "stuck"). Really you just need the
temperature ladder, which can be quickly and easily determined from the
T-REMD calculator website (and it's quite remarkably accurate).

With T-REMD, you are guaranteed to get an entire simulation done at the
desired temperature, with an equivalent amount of simulation done at each
of the higher temperatures. It's not the most efficient use of resources
probably, but it greatly simplifies setting up your simulation and
analyzing the results. And you can recapture some of that "wasted
computation" by reweighting the higher temperature structures.

The presumption here is that computers are becoming ever-more parallel
while the parallel scalability of our MD codes is not really improving that
much (since MD has even a theoretical limit to scalability due to the
inter-thread communication requirements). An easy way to address this is
to scale up with loosely-coupled replicas.

And then of course there are other REMD protocols that don't have *any*
wasted simulation, like constant pH-REMD or many forms of Hamiltonian-REMD
(like TI-REMD, REFEP, or umbrella sampling REMD).

In comparison to other expanded ensemble-like techniques, REMD is easy to
code, easy to set up, easy to analyze, easy to expand, and easy to

All the best,

Jason M. Swails
Rutgers University
Postdoctoral Researcher

This archive was generated by hypermail 2.1.6 : Tue Dec 27 2016 - 23:21:45 CST