[gmx-users] Force Fields and Electrostatics

[Michael Brunsteiner]

--- Eric Jakobsson <jake at ncsa.uiuc.edu> wrote:
> We have compared cutoffs with Ewald, cutoffs of different lengths with each 
> other, and PME with exact Ewald.  Also looked at different lengths vdw
> cutoffs.
> 
> The summary results that have guided us:

Interesting results, I've embedded a couple of questions and comments ...

> 1.  If you use cut-offs and have either ions of non-neutral charge groups 
> in the system, you will see definite artefacts at the cut-off distances in 
> the radial distribution functions.  If you have a system that you can parse 
> completely into neutral charge groups, you will not see those artifacts.

I guess that's the very reason why charge-groups are used in many 
force-fields, isn't it ? It is well known by now that, when working 
with charged groups/ions ewald summation (or any equivalent) is 
indispensable, but the question here is: For systems without ions/charged 
groups, has ever anybody done a systematic investigation to which extent 
the - rather arbitrariy - definition of these neutral charge groups can 
cause artefacts (as compared to Ewald ?)

> 2. If you use cut-offs in a lipid bilayer in water simulation, and you use 
> the computed surface dipole potential as a measure of whether the 
> electrostatics is good enough, you should use 20 angstrom cut-offs.  There 
> is a significant difference between the dipole potential calculated at 15 
> angstrom and 20 angstrom cut-offs, but not between 20 angstroms and 25 
> angstroms.

you must be a lucky man, considering the enormous amount of computers
they must have given you to cope with such a cut-off ;) ... sorry
couldn't resist ... but seriously, here another question arises:
Wouldn't you expect PME to be much cheaper (in terms of CPU time) than 
using a 20 A cut-off, at the same time giving you (at least) the same 
preciceness ?
And a converged (precise) result is not automatically a correct 
(accurate) result, is it ?. As far as I know most common force fields 
such as Grom(o/ac)s, Charmm, OPLS and, to some extent, also Amber were 
parameterzed by fitting the parameters to reproduce various experimental 
data. Since this was done at a time when computer speed was an even 
sparser commodity than it is nowadays this fitting procedure normally 
involved very short cut-offs (<=10 A). What sort of errors do you
expect to get when you use a 20 A cut-off togehter with a FF that was
fitted with a 10 A cut-off ?

Many (if not most) classical computational studies are concerned
with the estimation of non-covalent binding-affinities. Here normally
relative energies (the ranking) is of much more interest than absolute
values, and here specific short range interactions are commonly assumed to 
be by far more inportant than long-range interactions ... I am sure there
are examples showing that this is not generally true, but considering
the fact that you cannot be sure to get more accurate results when
you extend your cut-off (the force-field issue, see above), why bother ?

> 3. Based on simulations of hexane (and analogy between hexane and the lipid 
> interior of membranes) we feel that one should use a 20 angstrom cut off 
> for vdw in membrane simulations.  (I know that some people use less, but if 
> you do the liquid hydrocarbon calculations, the results are flaky unless 
> you use a cut off that long. The vdw does fall off as the 6th power, 
> therefore faster than the electrostatics, but there is no shielding of the 
> vdw---all the long range interactions are attractive, and therefore have a 
> cumulative effect at larger distances than one would otherwise think.)

Estimation of dispersion interactions is generally a very messy business,
especially when it comes to molecule-surface interactions. The quantity
to look at here would be a Hamacker-constant but predictions, as well
as experimental results for this constant vary wildly, and if a simple 
summation of LJ-C6 terms, up to what-ever cut-off, gives you a correct 
(as opposed to precise) result I would call this a lucky coincidence.

> 4. The artefact to worry about for PME is that, since it is not strictly 
> conservative, the center of mass of the whole system may move.  We never 
> had to use center of mass corrections until we started using PME.  We know 
> that artefacts associated with uncorrected center of mass motion and PME 
> grow to dramatic size much more slowly if one updates the neighbor lists 
> every time step; the big problem seems to come when particles move across 
> the explicit/Fourier boundary and the system doesn't adjust the neighbor 
> list immediately.  If you want to compute something really amusing, 
> simulate a box of water using PME, leave about 5 time steps (10 fsec) 
> between each neighbor list update, and use the NPE ensemble.  Many sins are 
> coated over by coupling to a temperature bath!

The question here is: are these errors (sins) systematic errors
or not. If not then you simply need to simulate longer to get
reasonable averages, and considering the efficiency of PME this
would probably pay out. If the errors are indeed systematic I would
appreciate if you could point me to some literature there.

cheers,
Michael


__________________________________
Do you Yahoo!?
Yahoo! Small Business $15K Web Design Giveaway 
http://promotions.yahoo.com/design_giveaway/