On Fri, Jul 18, 2014 at 12:55 AM, Niels Geudens <Niels.Geudens.ugent.be>
wrote:
>
>
> What a nice expression: NaNdyland.. :)
> I used pmemd.sander or cuda (single GPU) for the simulation, all of
> them blew up. I will reparameterize the molecule and see what happens.
> If it still goes wrong, I will let you know!
>
It's not necessarily the parametrization. Common things to look for are
atoms with zero vdW parameters (mainly hydroxyl hydrogens in the Amber
force field) that can collapse into another atom. Atoms with zero mass
(that are not extra points fixed to a particular frame of reference)
(although this would result in an immediate failure). There are not too
may things that can give rise to NaN's, and simply rerunning antechamber is
unlikely to be productive (unless you didn't check your parameters at all
and a bunch of them are 0's because parmchk couldn't determine suitable
substitutes).
It's also possible that your lambda value is too small, and the potential
energy surface is flattened too much in some region (have you tried
different values of lambda?).
Another thought is to specify a value of lambda that is formally equivalent
to standard MD. According to the paper, setting lambda=1 should result in
a PES that is effectively unscaled (so I think you should get standard NVT
behavior), but still should traverse the scaledMD code. Of course it would
seem that scaledMD documentation is rather lacking in the Amber 14 manual
-- I see only mention that it is a new feature in pmemd, to say nothing of
how to actually use it. But see if you get crashes with standard MD as
frequently as you got crashes with scaled MD. Then try turning on scaledMD
and set the lambda value to 1 and make sure it still behaves like standard
NVT dynamics.
Then start changing that input to match the one that fails for you.
Hopefully that could help narrow down the input that's causing the
problems (if indeed it is an issue with an input parameter or a bug in the
code). That'll either help you determine what parameters you should change
or give the scaled MD code maintainers a hint about where to look for a bug.
Hope this helps,
Jason
--
Jason M. Swails
BioMaPS,
Rutgers University
Postdoctoral Researcher
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Jul 18 2014 - 02:30:03 PDT
