[gmx-developers] Re: New implementations of Parrinello-Rahman

[Michel Cuendet]

Hi,


>The constraints are applied to the positions after the rescaling.   
>It's most efficient to just rescale the center of mass positions of
>the molecules, and to use the molecular virial instead of the atomic
>virial, but that's not really supported in GROMACS right now since
>there is no separate list of molecules set up.  It might be possible
>to use the molecular virial for the water molecules (or whatever the
>solvent is), since they are already defined separately, and the atomic
>for the other molecules.   I've heard that there's a very small issue
>with this (heard through the grapevine from Martyna and Tuckerman - I
>think it's for any set of constraints, not just constraints updated
>this way, if anyone has the reference, let me know), but in my tests
>of energy conservation, it's extremely small, and there's no drift in
>the conserved quantity for many nanoseconds.

I think that when considering a lot of small molecules (like the solvent) the molecular virial and rescaling is the right way to go, as the molecular virial avoids the computation of the intra-molecular force contributions. These intra-molecular terms can give rise to large numbers which then add up to a small number, introducing large errors. 

The case of solute molecules is a bit different in my eyes (beyond the fact that they are not defined as molecules in gromacs). Take a big protein in the middle of the box, and try to do molecular scaling. The molecular rescaling is an homotecy centered, say, at the middle of the box, applied to the molecular centers of mass. If the center of mass of the protein is in the middle of the box, the protein will never move, nor inflate. So also the atoms close to the surface of the protein will not move. But the first water molecule next to the surface will experience a displacement proportional to its distance to the origin, which might be quite big. This effect will make the whole first solvation shell to move back and forth from the protein surface... 

And myabe then you would even have a nice energy conservation, but still a non-physical solvation. We should be careful not to just aim for energy conservation...

So I think the idea of having atomic rescaling for the solute and molecular rescaling for the solvent is good. Now it should be checked that an hybrid rescaling scheme is consistent on the theoretical level. We should be careful not to just aim for energy conservation... Maybe in the case above, you would even have a nice energy conservation, but still a non-physical solvation. Also, the scheme chosen should try to minimize the itegrator-induced phase space compression, for example.

A more philosophical question : What is really the compressibility of a protein ? It might as well be very low. In this case, one would have to consider rescaling the 
waters with respect to the surface of the protein, and not to the middle of the box. But this is SF...

Best,
M

-- 
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Michel Cuendet, Ph.D. student
Laboratory of Computational Biochemistry and Chemistry
Swiss Federal Institute of Technology in Lausanne (EPFL)
CH-1015 Lausanne						
Switzerland                         	Phone : +41 1 693 0324
lcbcpc21.epfl.ch/~mitch                        +41  1 633 4194
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++