Re: [AMBER] Estimate error for TI calculations

From: Jason Swails <>
Date: Wed, 31 Jul 2013 13:21:34 -0400

On Wed, Jul 31, 2013 at 1:07 PM, Kumar Sonu

> Hi,
> I have done TI using 5 point gaussian quadrature lambda values for my
> calculations. Now I am facing problem in calculating error along with
> binding energy. Is there way to calculate error??
> for binding energy, I am multiplying corresponding weight of each point
> with Avg. dV/dL value and finally summing up for total energy. But I have
> no idea to estimate error?

You can measure the statistical uncertainty of the measurement using MBAR,
an analysis that's even supported by sander's TI module (see the ifmbar
variable in sander and the script that postprocesses the
resulting mdout file).

If you google "pymbar" you should be able to find the pymbar implementation
from Shirts and Chodera (it's worth reading their paper, too).

You may be able to get a general idea of the size of the error by
bootstrapping the DV/DL values you pull from the mdout files N times and
computing the free energy for each bootstrapped subsample. Then you can
take the standard deviation of your N subsamples as a (overly?)
quick-and-dirty estimate of your statistical uncertainty (note, MBAR is far
more rigorous in this regard).

In each case, the data must be uncorrelated, or your uncertainty estimate
will be meaningless.

Other errors, like force field and sampling errors, are far harder to


Jason M. Swails
Rutgers University
Postdoctoral Researcher
AMBER mailing list
Received on Wed Jul 31 2013 - 10:30:03 PDT
Custom Search