> down, always within 10ps. I understand nobody should expect a perfect
> match. Next time, I'd suggest my students not to switch between sander
> and pmemd in the middle of run. This should solve their problems :).
Getting to the underlying question, do these small differences between
amber6, amber7 and pmemd represent a systematic error / flaw that implies
that one cannot mix and match executables, codes, differing numbers of
processors or architectures within a given "project" or "trajectory"?
We know that MD simulation is chaotic; there is a nice article by
Braxenthaler et al. (Moult) in Proteins ~1997 that discusses these issues.
As Bob eluded to, perhaps there could be many theses investigating the
floating point operations and hidden errors in MD codes. To verify
that these are real systematic errors (and not random fluctuation or
non-accumulating or non-significant differences in evolution of the
dynamics) a large series of systematic tests would have to be
performed. This would be inherently tricky due to the chaotic nature of
the dynamics.
My feeling, having done tests like this with different pressure couplings,
SHAKE tolerances, etc to probe energy conservation in AMBER suggests that
on the < 50 ns time scale, these "artifacts" are not largely significant.
Of course, the flying block of ice problem was also insignificant when we
were limited to the 1 ns time scale. What is not clear is what will
happen as we run for longer and longer times, approaching routine
microsecond time scales. I worry about SHAKE approximations, multiple
time steps algorirthms and potential for non-random errors in the
integration (like lose of atoms in the pairlist).
In the end however, What really matters to me is whether we can give
insight into the biological phenomena of interest. Sure, we have to look
out for problems and artifacts and overcome them as much as possible, yet
even with a distance dependent dielectric insight into protein structure
and dynamics can emerge. We are using a simple model of the real system;
the model is full of inherent inaccuracies and limitations, yet our
current generation of force fields and methods can fold up a protein.
Pretty amazing if you ask me...
--tom
Received on Wed Apr 05 2006 - 23:50:07 PDT
