Re: [AMBER-Developers] Looking for post-doc/progammer to work on polarizable force fields

From: Ross Walker <ross.rosswalker.co.uk>
Date: Fri, 16 Aug 2013 11:34:34 -0700

Hi Dave et al,

I think it would be good to get a discussion going on this and we should
probably have a plan in place for how AMBER should approach polarization
for discussion at the next AMBER workshop.

My opinion is that the first thing that needs to be done is to decide
which of the polarization models is the path to follow - it will be crazy
and near on impossible to try to support all of the different approaches
and so it is probably wise to select one now, but select it carefully
within a number of constraints.

Some points for discussion:

It should be a 'simple' extension to existing force fields that provides a
measurable improvement. Complexity is not necessarily a good thing. For
example Amoeba is, I believe, probably the worst choice that could be made
for addressing polarization for large and long timescale simulations. It
is complex to parameterize, slow (and always will be) and in a lot of
cases less accurate than existing fixed charge models. Certainly it has
the flexibility to be improved but it will always suffer from being too
complicated in my opinion - at the complexity of implementing and
parameterizing Amoeba one would probably be better looking at ways to
improve existing quantum methods and make those faster. So for
polarization in AMBER, I think it should be:

1) A simple extension to the existing force field and thus easy to add to
all variants of existing code, easily compatible with our current topology
files, coordinate files etc, compatible with as many features as possible
as easily as possible - this includes things like QM/MM, various periodic
boundary treatments, second derivatives for normal modes (I don't even
want to think about what the second derivatives would look like for
Amoeba) etc etc and be able to readily show an improvement over fixed
charge models.

2) It should be simple for the user to parameterize. A lot of MD these
days is done to look at behavior of ligands and so it needs to be simple
enough that one can easily parameterize for huge libraries of ligands. The
charge derivation (including the polarization) needs to be well
documented, fast and automated if it is expected that people will adopt
these new force fields. Antechamber needs to be easily able to process the
polarizable force field for all reasonable ligands. I believe that a good
part of the reason FF02 was not widely adopted and extended within AMBER
was because the tools were never made available to determine the charges -
it was a manual iterative method and never very well described. Our
automated parameter fitting tools will need to be updated to support
fitting polarizable force fields (antechamber for example) so ideally one
should be chosen that makes this extension simple - otherwise it will need
substantially more effort to make this effective.

See for example: http://dasher.wustl.edu/ponder/papers/jctc-7-3143-11.pdf
which outlines the approach for Amoeba. It is at least step wise but far
from automated at present.

3) It will need to be FAST from the outset. This means it needs to be
LOCAL. The polarization will need to be constrained to local affects such
that it can be computed in a massively parallel fashion. Iteration is bad
but can be dealt with if it can be done in a very local reference frame
and thus in parallel - e.g. shake. If the polarization is such that it
requires serialization of key steps in the force calculation and
integration - for example requiring some form of non local knowledge
during the iterations then it will be doomed to failure. At the end of the
day if it can't run effectively on 10^5 or more threads it will never take
off without custom hardware designed to support it. If the approach needs
complex matrix algebra, or god forbid a diagonalization it is toast from a
performance perspective.

4) It should be numerically stable and ideally not require excessive
precision in the evaluation of the equations. Hardware will likely
continue to move further and further away from effective double precision
support so we should keep this in mind. For example the angle term in the
amoeba force field, expanded to order 4 with some very small pre factors
could potentially cause all sorts of havoc unless precision is kept very
tight.

The last two are in my opinion the achilies heel of amoeba. It will never
compete in terms of speed with pairwise additive force fields on modern
hardware. I think to really work whatever is decided needs to be able to
get within a factor of 2 of performance of existing classical
non-polarizable force fields. If it doesn't then I think the impact will
be minimal since instead of being a more accurate replacement for
classical force fields it will be rather a 'more' approximate approach to
doing QM calculations. This is by no means an unworthy goal, just that I
think the goal right now should be to improve classical force fields
rather than trying to come up with a slightly faster QM competing method.

It is unfortunate that engineering considerations dominate choices in MD
rather than pure scientific considerations but something we cannot ignore.

So I think we need to take a hard look at all the available polarizable
models, how they work, and then rank them in terms of a series of criteria
relating to their ease of parameterization, ease of programming, potential
massively parallelizability and performance. Sure accuracy is a key metric
here, obviously there is no point devising a 'slower' polarizable model
that does worse than existing classical force fields, but I think it
should be a criteria amongst several, not the overriding factor.

Ultimately 'utility' is a subtle mix of accuracy and statistical sampling
so we need to do this in a way that tries to find the sweet spot.

My 3c.

All the best
Ross



On 8/16/13 5:57 AM, "David A Case" <case.biomaps.rutgers.edu> wrote:

>Many exicting developments in molecular simulation revolve around the use
>of
>polarizable (aka non-fixed-charge) force fields. Prominent examples
>inlcude
>the "ff12pol" efforts from the Amber force field consortium, the Amoeba
>force
>field from Ponder&Ren, and the Drude particle ideas from (among others)
>the
>CHARMM community.
>
>The Amber codes are not currently set up to be very flexible or efficient
>in
>using such force field descriptions. I'd like to promote discussion and
>some
>real work in allowing greater flexibility in what we do (so that Amber
>folks
>can contribute more easily to ff development), and to speed up our codes
>in
>this area.
>
>I have some funds available to support a post-doc or programmer. I also
>think
>that coordination of efforts in various labs would also be helpful. If
>you
>have ideas or comments, please post to the amber-developers list.
>Candidates
>for a job should contact me directly.
>
>...thx....dac
>
>
>_______________________________________________
>AMBER-Developers mailing list
>AMBER-Developers.ambermd.org
>http://lists.ambermd.org/mailman/listinfo/amber-developers



_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Aug 16 2013 - 12:00:03 PDT
Custom Search