Summary of replies about simulated annealing
Good day, all -
Some time ago I posed a question to the net about simulated annealing. Finally,
here is a summary of responses I received. Warning: it is somewhat lengthy.
Direct quotes are in parenthesis. Everything else is my edited summary. My
apologies if my editing loses some of the flavor (or, God forbid, the meaning)
of the responses. My thanks to all who offered opinions.
My original question follows:
>I am trying to refine some structural homology models of proteins. One
>approach I am considering is simulated annealing. However, I want to ensure
>that I understand the concept first. Please correct me if I am in error.
>As I understand it, simulated annealing is simply a molecular dynamics
>simulation where the protein is heated to a rather high temperature then
slowly
>cooled down to zero. The protein is then locked into some low energy
>conformation which it was close to at the higher temperature. In order to
get
>accurate results, one needs to perform the simulation several times and
obtain
>a consensus structure. How am I doing so far?
>Okay, so here are my questions.
>(1) Can the procedure be implemented using the standard MSI CHARMM package
or
>is a special (and maybe proprietary) version needed? Has anyone a script
file I
>can use to learn by example?
>(2) How high a temperature is needed? 300K, 600, 900, more?
>(3) Any good references for learning how to use simulated annealing?
>(4) Am I missing anything in my understanding of the technique?
*************************************************************************************
From: Bob Funchess
MOLECULAR SIMULATIONS HOTLINE SUPPORT
16 New England Executive Park
Burlington, MA 01803-5297
hotline()at()msi.com
"You are correct about simulated annealing. One thing is that you need
to go to a much higher temperature such as 5000K."
He also provided a much appreciated CHARMm script as an example for performing
simulated annealing. I will be happy to provide the script to anyone upon
request.
*************************************************************************************
From "arne" (arne()at()mango.mef.ki.se)
"I think that you are correct in your assumptions."
Some other comments are:
1. "I am quite doubtfull that a sim. annealing approach
will work without any "extra" information, such as
NOE constrains or something else."
2. "Temperature:
I am using a 500 K temperature without any cooling
period for modelling BPTI and small helical segments."
3. "references:
The original paper was :
Brunger, Clore et al, PNAS,83,3801-3805 (1986)"
*************************************************************************************
From: David C. Doherty
Computational Scientist
Minnesota Supercomputer Center
doherty()at()msc.edu
"Functionally, this is correct, but usually simulated annealing involves
taking monte carlo steps, and following a schedule of temperature lowering.
How you design that schedule (how many steps of MC or MD to take) is something
of an art."
Temperature is suggested to be of 700-1000K to start.
*************************************************************************************
From: Georgia B. McGaughey
Computational Center for Molecular Structure and Design
Department of Chemistry
University of Georgia, Athens, GA 30602
georgia()at()Huckel.chem.uga.edu
Georgia provided two references for the use of SA with NMR data.
1. James TL. Relaxation Matrix Analysis of Two-Dimensional
Nuclear Overhauser Effect Spectra. "Current Opinion in
Structural Biology, 1, 1042 (1991).
2. Clore GM, Nilges M, Gronenborn A. Determination of Three-
Dimensional Structures of Proteins in Solution by Dynamical
Simulated Annealing with Interproton Distances Derived from
Nuclear Magnetic Resonance Spectroscopy. Computer-Aided
Molecular Design, Richard WG, ed. 203 (1989).
*************************************************************************************
From: Emil Marcus
Biophysics PhD Student
University at Buffalo / Roswell Cancer Institute
marcus()at()acsu.buffalo.edu
"Simulated Annealing is a Monte Carlo ("random draw",
"throwing the
dice") method that can be quite successful in dealing with large
combinatorial optimization problems, such as finding the conformation
of a protein (i.e. find the position - cartesian coordinates or
dihedral angles of all, or only a targeted number of atoms, such as
those which are part of the backbone or side chain). The idea
underlying most conformational searches is to find a set of
coordinates (cartesian or dihedral) that minimize the potential
energy of the system. Unfortunately, the conformational space of a
protein features (very)many local energy minima... Simulated
Annealing overcomes the major inconvenient of a typical
gradient(derivative) - based method: that of remaining/being trapped
in a local minima. SA *samples* the conformational space, prevents
being trapped in local energy minima because, due to its Monte Carlo
nature it is able to "jump" from one minima to another. It does
not
promis/guarantee the global minimum, however, if used carefully, and
in combination with other search methods (such as homology...) can
provide good results: consider for instance the pentapeptides,
Met- and Leu- Enkephalin."
"No SA option is provided (even) in the last version of Charmm,
Charmm - Version 22.0.b - April, 1991, I have recently used.
Neither is it available with the Amber or Ecepp force fields."
*************************************************************************************
From: Lee Herman
HERMAN()at()ULNA.BWH.HARVARD.EDU
"The heating phase is a trick to push the structure into parts of the
conformational surface that are dissimilar to the starting structure.
This structure is then allowed to minimize (cool) to a local minimum.
This final structure can be quite different locally from the "hot"
sampled structure, allow in terms of tertiary structure you're right,
they're probably pretty close.
"Watch out when it comes to sampling conformational space "a few
times".
The degrees of freedom for a structure such as protein can be enormous.
Adequate sampling is essential in searching for a "global minimum" (if
you think it's important, that is. It may not be.). Adequate sampling is
also necessary in order to properly estimate properties. Find literature
on how many samples are appropriate for the kind of system you're looking
at."
*************************************************************************************
From: Gerard J. Kleywegt
Department of Molecular Biology
Biomedical Centre
Sweden
rd()at()xray.bmc.uu.se
"We use this technique often while refining protein structures against
crystallographic data. In those cases, we start at 3000 or even 4000 K
and do a fast-cool (steps of 50 K) or a slow-cool (steps of 25 K) down
to 300 K. I don't know about CHARMm, but XPLOR (which, I think, is also
sold by MSI) does a good job at it."
He strongly recommended using some experimental data as constraints in the
SA.
*************************************************************************************
From: Chris Chen
Agouron Pharmaceuticals, Inc
San Diego, CA 92121
chen()at()tbone.agouron.com
"The higher the temperature, the better the calculation theoretically. 900K
is a best choose when I used MMOD to do the calcs. Another thing probably
need to be mentioned. When you cool down the very high temperature, you may
have to consider which way is better: stepwise annealing or contiuns
annealing. That will cause different results."
*************************************************************************************
From: Jim Poole,
Ohio University,
Dept. of Chemistry
jim()at()quanta.phy.ohiou.edu or jpoole()at()oucsace.cs.ohiou.edu
"You have probably already gotten lots of response but here are a few
references that I have found for sim.-ann.
Treutlein, et al. (1992) Biochemistry 31,12726-12733
best Nilges and Brunger (1993) Proteins:Str. Fxn. Gen. 15, 133-146
Brunger (1991) Ann. Rev. Phys. Chem. 42, 197-233
Kirkpatrick, et al. (1983) Science 220, 671-680"