> I'm not claiming to be elite, but I am a big fan of being exceedingly
> picky about these things. So, I vote to make them all the same and
> clean up the tests. In fact, go to NIST and make sure they are the
> official, most recent, umpteen-decimal versions. I also vote to make
> all the functions use the same constants from one central location
> unless there is some very good reason a certain function needs it to be
> different.
I do not want to make a mountain out of an anthill, but I note that we
have to be careful as presumably the simulation results depend on the
choice of constants. Changes will effect comparisons to other codes and
the definition of the "gold standard".
To date we have not held to a gold standard; in fact, the current lead
(i.e. Pb) standard for a particular force field that we have applied
to-date is code snapshot, run-time options, methods, and implementations
used *at* the time the force field was developed.
To get CHARMM and AMBER forces to match to six decimals of precision
required altering conversion factors / constants (in the two codes)...
The most significant is the "charge" to kcal/mol change which is different
in both AMBER and CHARMM (and both are wrong). The definition of "pi"
also affected the results. Which code is correct? Or are the results
with both simulation codes correct or both incorrect?
By moving towards "the correct" or gold standard, this has to be clearly
documented, and to my mind, validated and assessed in terms of the
implications of the changes. For example, is it still the Cornell et al.
force field when long-range vdw corrections are applied (altering the
system density and therefore running with vdw parameters that were not
optimized with long-range corrections), when different PME approximations
are used, with different approximations to the transcedentals, with
different code "constants"?
Are "GROMACS" implementations of the AMBER force fields really getting
equivalent results to "AMBER" when they run fast and loose with single
precision, transcedental approximations, questionable time steps, etc?
This is a much larger question than simply altering a few force constants
by a small amount.
So, the question is do we need to dig deeper towards accuracy or assume
that AMBER, as implemented, is AMBER and therefore correct. As we change
the constants, are the old results now incorrect? In most cases, the
changes will likely be small and likely not lead to systematic
accumulation of (larger) error. On the other hand, altering how the 1-4's
feel each other in terms of small changes in charge could in turn
influence dihedral barriers; or small changes in vdw/charges could alter
TIP3P diffusion or ionic solvation free energies, etc, etc...
As changes are made make them clear so that both our users and other codes
can understand differences in results and attempt to remain "compatible".
Note that coming up with a gold standard of benchmarking "results" has
been something we have talked about for a long time and really should get
to to understand the implications of change.
--tom
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Mon Oct 11 2010 - 12:30:04 PDT
