Re: [AMBER] About T-REMD convergence

From: Jason Swails <>
Date: Mon, 16 Sep 2013 06:31:59 -0400

On Mon, Sep 16, 2013 at 5:17 AM, Francesco Pietra <>wrote:

> As to
> From: Carlos Simmerling
> <<
> >>
> Date: Tue, 18 May 2010 06:54:14 -0400
> Much depends on the initial structures. If all the same, you realy
> need a second run from different coordinates. If different, the data
> from all replicas should match. This means that you can extract the
> temperature data from each replica and they should match. For example,
> fraction folded at 300k must be the same for replica 1 and 2. This
> takes some effort to analyze.
> I assume that for a 32-replicas T-remd in the 314-600K range, the above
> criterion should be applied at replicas sorted at 314K, if the interest is
> in the system at 314K. Replicas extracted at all other temperatures should
> simply be neglected?

I think you misunderstand what Carlos said. A good test of convergence
for REMD simulations is to analyze each replica (NOT each temperature), and
make sure that each replica contains the same information as every other
replica (especially the low-temperature sub-ensemble). In 'good' REMD
simulations, each replica will visit each temperature approximately an
equal amount of time. If you analyze all of the 314K snapshots from each
replica, the quantities that you compute (e.g., RMSD distributions, RMS
fluctuations, average structure, percent folded, etc.) for each replica
should be equivalent. Technically the same is true for all other
temperatures, but since high temperatures yield far more available states
than low temperatures, those are harder to 'converge'.

The above is just a test for convergence. When you actually want to
analyze the data, typically people just extract the ensemble from the
temperature they're interested in and analyze that ensemble. While this
throws away all of the information generated at higher temperatures, that
information is less useful, anyway. Force fields are not validated at high
temperatures, so the results cannot really be trusted like they can at the
temperatures that the force fields were parametrized. However, there are
techniques you can use to 'reweight' the high-temperature data to
supplement the information you have at low temperatures. A good example is
the multistate Bennett Acceptance Ratio. [1,2]

> I started such a T-remd under GB conditions (abandoning implicit water
> after Prof Simmerling 2013 paper) and progressively increasing temperature
> for a 34aa peptide under restraining of dihedrals for a short initial
> stretch (the only portion diffracting enough under X-ray).

I'm confused here. GB is implicit solvent (which you claim to be

> Debug T-remd
> with 32 replicas and 3700 steps for each replica at ts=0.2 fs (exchange
> ratio higher than 0.7), thus all rigid bonds, show better folding at 600K
> than 314K. Although the conformation is unknown, there is a more ordered
> organization ant 600 than 314K. Is that acceptable in order to go to
> production?

If you start from the same snapshot for each replica, 3700 steps is really
not enough to ensure that the acceptance ratio is not artificially high
since the structures at each temperature may be quite similar to each
other. If, however, the structures are all different, a 70% success rate
means you are using too many replicas.

> Good exchange (ratio 0.4) also with 16 replicas, however with very little
> gain of computer time as I am bound to use 64 nodes.

Is the interconnect too poor to use 4 nodes per replica? GB typically
scales quite well

All rigid bonds is something that for MD in general I dislike, particularly
> if moving ligands are under scrutiny. What about for such a type of T-remd?

Why are you making all bonds rigid? I would advise against it, since that
is atypical for Amber force fields.


[1] (defines
[2] (MBAR with
temperature re-weighting)

Jason M. Swails
Rutgers University
Postdoctoral Researcher
AMBER mailing list
Received on Mon Sep 16 2013 - 04:00:02 PDT
Custom Search