[gmx-users] FEP softcore problematics

David Mobley dmobley at gmail.com
Wed Jun 21 17:00:19 CEST 2006


The paper I'm talking about is by Anwar and Heyes and is the first
entry here: http://scholar.google.com/scholar?hl=en&lr=&q=anwar+heyes+charge&btnG=Search

> Thank you for the detailed answer.
> It sounds reasonable. :)
> For sure it's the best way to make the calculations separatly and sum up
> the energies at the end. You mentioned, that, in the case of splitting,
> the "density" of snapshot samplings at lambda=0,1 can be lower.
> So I guess, one reaches approx. 30-35 samplings, when splitting the
> terms. In the first place, this is no big deal, because of the
> possibility of parallelization, but for screening purposes in the
> future, the amount of simulation time in total is quite critical for me.

I agree with Berk that it is more efficient to do it in three steps
than to try to do it in one step. By "more efficient", I mean you can
get by with fewer total lambda values. Electrostatics are usually
smooth enough you can change electrostatics alone using only 5 lambda
values. I can usually do vdW alone with 10-16 lambda values, so that
leaves you with a total number of lambda values for the three steps
combined of somewhere between 20-26. You may be able to get by with
fewer depending on the accuracy you want. I doubt you can get by with
this few (and still get the right answer) for any scheme where you
change both the charges and electrostatics at once, unless the Anwar
and Heyes scheme works as well as they think. I doubt the
naphtalene->naphtalene test in the paper is rigorous enough to know
for sure whether this is the case.

I suggest before getting too carried away you at least *try* doing the
transformation with a three step approach and see how it compares in
terms of efficiency. Also, if you do the three step approach
correctly, there should be no charge-charge problems, so you should be
guaranteed to get the right answer if you run long enough, which is
more than you can say for the one-step approach; it will give you a
baseline to compare to when you try other things.

There are several different three step approaches depending on what
exactly you were doing. Berk's is one. An alternative one would be:
1) Neutralize any atoms which are being disappeared
2) Modify LJ interactions from A state to B state
3) Change charges to final charges for B state.

Again, the main idea is that you just want to avoid ever having charge
centers that have LJ parameters of 0, or LJ parameters which are going
to zero. As long as you avoid this, you can pretty much do it however
you want. Also, it is FAR easier to change electrostatics interactions
than LJ parameters.


More information about the gromacs.org_gmx-users mailing list