Date: Thu, 17 Oct 2002 14:08:16 -0700
From: "David A. Case"
Subject: cutoffs in GB

On Thu, Oct 17, 2002, Andreas Svrcek-Seiler wrote:
>
> >
> > I used cutoff=16.0, igb=2. The results fit very well with all of the
> > experimental data. I read sometime ago that cutoff in GB is recommended to
> > be set as a large value. I guess this is a personal opinion instead of a
> > general recommendation to all amber users. Correct me if I am wrong.
> ... A hard cutoff is generally a bad thing - in principle.

> When you use a cutoff in conjunction with GB, you're actually cutting off
> two things:
> 1) The contribution of atom j to the (inverse) effective born radius of
> atom i if they are seperated by more than than 'cutoff'.
> This is rather harmless, since this contribution drops of with
> dist**(-4) and the radii are not too accurate anyway (crudely
> speaking, with lots of details omitted).
> *But* this cutoff alone still generates discontinuities in the energy
> landscape.
>
> 2) The Coulomb (a), 'pairwise solvation' (b) and Van der Waals
> anergy and force contributions (c).
> This is worse (produces 'larger' discontinuities) for (a) and (b)
> although they *tend* to cancel each other (just look at the GB
> and Coulomb formulas).
> (c) is harmless at 16 A (except for purists, maybe).
>
> I have also seen evidence that a cutoff is not worse than no cutoff
> in implicit solvent MD (when comparing 'simulated' and experimental
> structures), as long as the heat bath takes care
> of the drift induced by the cutoff. However, doing so one relies
> on the cancellation and/or masking of individual errors,
> which should at least always be kept in mind.

Andreas gives a really nice answer. Let me just add a couple of points:

(a) A 16 Ang. cutoff in GB is less severe than with regular simulations,
since the effective potential is well-screened at those distances, i.e.
the Coulomb + polarization potential is something like 1/80 of what it
would be in vacuum. (This is just a restatement of the "tend to cancel"
wording above.

(b) Things are even better with non-zero salt concentrations, since the
effective potentials at large distances fall often even more quickly.

(c) One thing that can help is to use nrespa=2 or even 4 for GB simulations;
this will reduce the time spent on computing long-range interactions, while
keeping their effects in the simulation. We are still in the process of
trying to figure out and incorporate the best multi-time step algorithms.

I think for many problems, you'd be hard-pressed to distinguish a 16 Ang.
cutoff simulation from one with no cutoffs, but there is no guarantee that
that will always be true.