From: Hannes Loeffler <>
Date: Thu, 29 Jan 2015 11:05:46 +0000

On Wed, 28 Jan 2015 17:25:02 -0500
Jason Swails <> wrote:

> > On Jan 28, 2015, at 3:52 PM, wrote:
> >
> > I don't quite understand your second paragraph about computing the
> > averages.
> H = lambda*H1 + (1-lambda)*H2 <-- H is effectively a lambda-weighted
> average of the two end-point Hamiltonians (at least in this flavor of
> TI, which is the one implemented in Amber). The naive, easy way to
> do this is compute H1 and H2 separately and then combine the energies
> and forces in that lambda-weighted average. This is what sander
> does, and it hijacks multisander to do it.
> For the terms that don’t depend explicitly on lambda (and by
> extension don’t contribute to dV/dl), sander still computes them in
> both states and averages them. By contrast, pmemd computes the
> lambda-independent parts only once and averages the lambda-dependent
> parts. This results in a nice speedup for PME by roughly halving the
> expensive direct-space sum (but not much for GB given that the most
> expensive part is almost entirely lambda-dependent).

Ok, many thanks. I have read Joe Kaus' paper on this and understand now
what you have said earlier.

> > I am not sure if there is currently sufficient evidence to say that
> > MBAR is more efficient than TI or vice versa.
> Well I think it’s pretty clear that MBAR with TI is more efficient
> (in its worst case, it is just as efficient). Especially since all
> of the MBAR energies can be computed with virtually no cost by sander
> and pmemd, I think the only real barrier to incorporating MBAR as a
> routine part of every TI calculation is the lack of mature tools that
> automate it.
> As for a non-TI method using MBAR to compute energies, I’d be
> skeptical of claims that TI is worse...

Hm, what other methods to you mean? BAR/MBAR and also those
overlap methods are in essence all FEP methods but in contrast to the
Zwanzig forumla you need to explicitly sample the energies at both
end-states (or several such "pairs") and then plug the data into the
respective formula. So it doesn't really matter what exact method you
used. TI is one possibility as the energies for any set of the other
lambdas can be easily computed as you say. Or do you have something
more sophisticated in mind which combines the TI gradients with MBAR?

The only tool for MBAR I am aware of is pymbar.


AMBER mailing list
Received on Thu Jan 29 2015 - 03:30:03 PST
Custom Search