AMBER Archive (2005)

Subject: RE: AMBER: sander segmentation fault

From: Ross Walker (ross_at_rosswalker.co.uk)
Date: Wed Feb 02 2005 - 20:05:41 CST


Dear Kristina

In my experience..........

> larger than maxcyc for minimizations involving rigid water
> models like TIP3P and
> SPC/E, because conjugate gradient minimization on these rigid
> models often
> resluted in LINEMIN failures. Looking through the current
> version of the DNA

Steepest descent can be very slow to converge so it is often wise to also
use a few conjugate gradient steps. Typically steepest descent will not
suffer from linmin failures as it just follows the force straight down hill.
Conjugate gradient is much more sensitive and will often suffer from linmin
failures if you either try to minimise with shake turned on OR try to
minimise too far.

The first scenario is what you are refering to above. Namely that when you
use rigid water models you should have shake turned on. This will generally
break conjugate gradient minimisation. The option here is to run more steps
of steepest descent or turn off shake. Since the purpose of the minimisation
is simply to remove any bad contacts, to stop the MD blowing up, rather than
to locate an actual minimum switching shake off during the minimisation
stage does not cause us a problem (as long as you turn it back on for MD).
In this way we can do less steps of minimisation than we would need to with
just steepest descent. This really is a matter of personal choice though.
Really all you are doing minimisation for is to remove any large forces
before MD. So, if one has a good starting structure, a few hundred steps or
so of steepest descent should be good enough. My aim in the DNA tutorial,
however, was to get people in the mind that they should try to minimise
their system more than is necessary before starting MD.

With regards to the second point, conjugate gradient minimisation will
generally break once your system gets very close to a minimum. If you do
minimisation without shake, and your system is not hopelessly bad, then this
should be the only reason you typically see linmin failures. In order to get
your system closer to the minimum it is then necessary to switch to more
robust second derivative methods like Newton Raphson. For MD purposes such
'aggressive' minimisation is not required. Hence this type of minimiser is
not implemented in Sander. You should check out nmode if you want to get
closer to a minimum.

I hope this makes sense.
All the best
Ross

/\
\/
|\oss Walker

| Department of Molecular Biology TPC15 |
| The Scripps Research Institute |
| Tel:- +1 858 784 8889 | EMail:- ross_at_rosswalker.co.uk |
| http://www.rosswalker.co.uk/ | PGP Key available on request |

Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
 

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu