Amsol performance
A couple of weeks ago, I wrote:
>>>>
I am running Amsol 4.5 on an Indigo2 (150Mhz), trying to do solvation
energies of compounds with ca. 15 heavy atoms. These systems optimize for
the gas phase in about 2 cpu minutes using the keyword "AM1". For the
solvation, I am using "AM1 SM2" and 4 cpu _hours_ later, things are
still
chugging along. I have no experience with the solvation models, and would
like know what order of magnitude of cpu times I should be expecting. Also
, are there keyword combinations which will reduce this (I haven't tinkered
with e.g. PULAY, since the gas phase calculation went so quickly)?
>>>>
Here are the interesting replies. Note particularly Cramer's remarks. I have
also appended a summary of my experience.
****
I have used Amsol on an SGI 380 running as single user - my cpds were
peptides - 6 residues - and Amsol AM1 SM2 combination run for 1 week
without convergence. This was with Amsol 3, to be correct. Using
Amsol 4.0 on Ala-Ala took a several hrs on a R4400 Challenge M
Generally, it is known to take a long time to converge.
Authors of Ampac 5.0 (Semichem - Andy Holder) claim their implementation
of Amsol is much faster. Never tried it, though.
One suggestion made by Christopher Cramer was to do AM1 SM2 1SCF on the
gas optimized cpd - to have an idea of solvating that conformer.
Tudor
****************************************************************************
* Tudor I. Oprea, MD PhD Tel: (505) 667 2682 *
* Postdoctoral Research Associate Fax: (505) 665 3493 *
* Theoretical Biology and Biophysics (T-10) Email: *
* Los Alamos National Laboratory tudor ^at^ t10.lanl.gov *
* Mail Stop K710, Los Alamos NM 87545 *
****************************************************************************
*****
I observed similarly long run-times in my limited experience with AMSOL.
If it's not too much trouble, would you please forward any useful replies
you receive to: pearlman ^at^ vax.phr.utexas.edu ?
Thanks a lot.
-- Bob Pearlman
*****
I feel I can answer this question fairly authoritatively . . .
The SMx solvation models as presently designed do not allow for analytical
first derivatives. We are working on this, but for the moment the only way to
do an optimization is by brute force finite differences. (Actually, even the
gas-phase optimizations are not truly analytical, but they get away with a
trick at the NDDO level that fails to be useful for solvation).
While this might sound depressing, there is a flip side to the coin.
As a rule, reoptimization in solution adds less than 5% to the total
solvation free energy relative to that calculated for the gas-phase geometry
(the only exception is for very polarizable systems, where small geometrical
changes can allow significant distortion of the electronic structure). Many
reports in the literature have appeared to this effect by other authors using
our models, and we have made these observations too (I'll send you refs if
you want, but don't want to waste bandwidth on the net). Since that 5% is
typically smaller than the intrinsic error in the model, it is negligible.
The frozen geometry calculations only take a few cpu seconds for a system of
the size you mentioned, so it is trivial to do the 1SCF calculations, and
then decide if you want to invest more time in the full optimization.
I add as a final caveat that we are always skeptical about semiempirical
geometries after having found time and again some fairly dramatic error
in structure (especially intramolecular hydrogen bonds, five-membered rings,
amides, and transition states that can exhibit biradicaloid character). Our
preferred approach these days is to optimize at a more trustworthy level
(depending on your system, that may be anything from a force-field to
converged QM calculations) and then add solvation from an SMx calculation at
that frozen geometry.
In closing, there are electrostatics-only continuum models that permit
analytical energy derivatives (e.g., the rather crude Onsager model available
in G92 and more refined versions that I believe are slated for release
sometime relatively soon, also Tomasi's PCM approach or Rivail's multipole
expansion in idealized cavities, coded into GAMESS-UK and the former perhaps
in MONSTERGAUSS). If you are dealing with molecules that do not have
stationary gas-phase structures (e.g., an amino acid zwitterion) then you
might consider such an approach. The virtue of the SMx models, however,
is that they have been developed to simultaneously account for
non-electrostatic solvation effects, and thus predict solvation free
energies that may be compared directly to experiment (Version 4.5 contains
only water models, soon-to-be-released version 5.0 also contains
hydrocarbon models, other solvents are in various stages of current
development).
Best regards,
Chris
P.S. You'll find the above summarized into one sentence in section 8 of the
users' manual (which illustrates a nameless universal truth . . . )
--
Christopher J. Cramer
University of Minnesota
Department of Chemistry
207 Pleasant St. SE
Minneapolis, MN 55455-0431
(612) 624-0859
cramer ^at^ maroon.tc.umn.edu
A later exchange between Cramer and myself:
> I'm prepared for
> significant time penalties for a 2 minute run (1+ order of magnitude), but
> the 2-3 that we seem to be seeing is borderline acceptable. We feel this
> is something needed to do, though, since we are (among other things)
> comparing the method to COSMO, which is fast if nothing else.
>
COSMO is a pretty method, although I would say it suffers from (i)
unoptimized radii for defining the dielectric boundary and (ii) it is, of
course, a pure electrostatic model and unable to handle the
non-electrostatics. Andreas Klamt was through here recently, and we had some
interesting discussions on some of these issues. We also compared the two
programs, and decided that the generalized Born solution to the Poisson
equation, and his Green's function approach, give much the same answer for
equivalent atomic radii.
Incidentally, Don and I wrote a review of continuum methods for Rev. Comp.
Chem. and, if the damn volume ever comes out (it's only been 17 months now)
we made a fairly large comparison between COSMO and various other methods,
including our own. Bottom line was that COSMO seemed to slightly overpolarize
things, in our opinion, but otherwise correlated nicely with SM2 solvation
free energies. If you'd like, I'll send you a preprint of that chapter.
> Thanks for the information. My first job ran about 5.5 cpu hours, but the
> second (an amide. . .) is 15 and counting.
>
In the superstitious tricks that you might find useful category, it often
turns out that a small perturbation from the gas-phase geometry (say 0.1
angstroms in one bond length) helps speed convergence. That is because the
default optimizer in AMPAC is sometimes a bit confused when it is started out
very near a minimum-energy geometry. It thrashes around a lot before deciding
it was in a pretty good spot to begin with. On the other hand, if it has at
least one decent gradient to minimize, it seems to consider itself to have
converged more effectively. If you can spare the time, you might want to try
such an approach on your recalcitrant amide and see if it helps.
Best regards,
Chris
--
Christopher J. Cramer
University of Minnesota
Department of Chemistry
207 Pleasant St. SE
Minneapolis, MN 55455-0431
(612) 624-0859
cramer ^at^ maroon.tc.umn.edu
*****
Now, my experience:
I have made 20 runs with molecules ranging from 16 to 24 heavy atoms. I have a
range of cpu times from 1.3 to 20 hrs on an 150 MHz Indigo2, with a mean of 7-8.
Of the runs, 4 were longer than 15 hours, 8 were 5 hours or less. I don't know
how "fair" this test is or how "typical" my molecules are --
although they are
quite typical for ag chem ;-) -- as Cramer comments, there did appear to be
some thrashing about but I did not see any correlation in run-time with atom
count, specific functional groups, etc.
I haven't checked the values for 1SCF solvations of the gas phase structures
for most of the compounds, but if the values are acceptable, this will be a
blessing: the run time is only on the order of a minute.
sb
===========================================================================
Stephen B. Bowlus, Ph.D. Computer-Aided Molecular Design
Research Division
e-mail: bowlus ^at^ sandoz.com Sandoz Agro, Inc.
Phone: + 1 415 354 3904 975 California Ave.
Fax: + 1 415 857 1125 Palo Alto, CA 94304
===========================================================================