[gmx-users] Implicit solvent

[Erik Lindahl]

Andrew Horsfield wrote:

> Hi,
>
> I am interested in finding solvation energies using an implicit solvent so
> that I can use geometry optimization instead of MD (to save time). As far
> as I can see, Gromacs contains several components needed to do this
> (solvent accessible surface calculation, electrostatic solver), but cannot
> yet perform the full calculation.
>
> 1) Has anyone used Gromacs for implicit solvent calculations?
>
> 2) If so, how did you do it?
>
> 3) If not, is anyone working on introducing this functionality?
>
> Cheers,
>
> Andrew

Hi  Andrew,

I'm working on it, although it will not be in 3.1.

This is actually quite complicated if you want to do it accurately. The
problem is that you need the surface derivative, but this entity has
discontinuities when two atom spheres collide or separate. This is partly the
reason why a lot of people have experienced random crashes with implicit
solvent models.

Now, the good news is that we've probably solved this, but it's kind of
complicated and involved triangulating the whole protein. The theory is ready
and coded, and we're only trying to make it faster right now. This will be a
"real" and accurate implicit solvent engine.

I've also been talking with the folding at home people here at Stanford, and
since they're interested in using it for their simulations we might also
include their simpler surface calculation.


The main problem is actually the speed of the surface calculation! If you
look at some of the published data they usually get 3-5 times speedup with
implicit solvent, even though the number of atoms is only a fraction of that
of a water system. This means the surface calculation probably takes 50% of
your runtime or so. Since the MD part of Gromacs is usually quite fast, you'd
end up with the surface calculation taking 90% of your time and a simulation
that's only a factor two or so faster than the full explicit solvent
alternative (This is a gross generalization, and not always true, but you get
the general idea :-).

So, I *really* want implicit solvent, but the algorithms need some
optimization to be effective - that's what takes time!

Cheers,

Erik