[gmx-developers] Re: minimization limitations

[Mark Abraham]

Igor Leontyev wrote:
>>> The bug with minimization seems to be fixed now, see
>>> http://bugzilla.gromacs.org/show_bug.cgi?id=332. Thank you Berk. With
>>> corrected sources "csettle.c" and "minimize.c" the new 4.0.5 version is
>>> able
>>> to minimize my MD system at least on the same level as it was 
>>> possible in
>>> 3.3.1 version.
>>>
>>> At this point I would like to mention 2 more limitation of the
>>> minimization
>>> procedure implemented in gromacs.
>>> 1) Since a dynamics of complex biological system depends on quality of
>>> the
>>> minimized structure, a fine minimization is very important.
> 
>> Generally, I would think not. The process of assigning atomic velocities
>> randomly sampled from a suitable distribution "undoes" the finest part of
>> the minimization. Then one needs to equilibrate, possibly under position
>> restraints, which also leaves the structure perturbed from the input
>> structure, but hopefully still in the same local minimum region. See also
>> comments in manual section 3.10.
> 
> It's true if the finer energy minimization do not exceed ~(kT) 
> relatively to
> the "steep" method for each site of the complex protein+solvent system. But
> it's hard to ensure for each site. Moreover, there are applications of the
> minimized structures other than MD simulations, in which there is no atomic
> velocities.  For example, continuum electrostatic pKa calculation needs as
> an input the protein structure with optimized hydrogen positions.

This example was excluded by your earlier description of "dynamics of a 
complex biological system", but anyway for such a pKa calculation, 
optimizing such protein hydrogens with respect to some (arbitrary?) 
frozen configuration of waters doesn't seem likely to be better than 
optimizing with respect to some flexible configuration of waters, or 
even without waters. Averaging over several such flexible-waters optima 
might be better still.

>>> Is there way in
>>> gmx to reFINE minimization of the system with many different constraints
>>> ?
> 
>> See manual. You can make the waters flexible, or use L-BFGS for
>> single-processor minimization.
> 
> 
> I have been tried to use all described in manual options but result was not
> satisfactory.
> Flexible water model does not solve the problem for my system with
> constrained hbonds and other constraints or frozen atoms. L-bfgs results to
> the structure where some hydrogens are locked in the local minimum which is
> far away from the global one.

Well maybe you need simulated annealing.

> I just ran l-bfgs minimization starting from
> the minimized by the "steep" method configuration and found that one of
> van-der-Waals-less hydrogen atoms in the final structure is transferred to
> the vicinity (~0.1A) of some negative nitrogen but far away (~3.0A) from
> its heavy atom . This local minimum is far away from global one due to huge
> bond-stretching penalty of the h-bond.

It moved because it was relaxing some other more severe strain, created 
either by the structure or the unbalanced model physics created by the 
constraints. What kind of non-water hydrogens should be modelled without 
  vdW?

>>> 2) After steepest descent minimization  of the protein+solvent system,
>>> the
>>> positions of the frozen TIP3P water molecules are slightly shifted (~0.3
>>> A).
>>> The shift of frozen molecules is really bothering.
> 
>> Are you centering the structures on the same frozen atom? Otherwise you
>> may just be seeing a shift of reference frame.
> 
> "Frozen" (by freezegrps) means in respect to the initial configuration,
> isn't it?

Dunno. Presumably so.

Mark