From: Aron Broom (broomsday_at_gmail.com)
Date: Mon Mar 25 2013 - 16:43:54 CDT
I think Niklaus' synopsis is excellent. I would say though that I think
metadynamics and ABF are fairly similar in their exploration. Metadynamics
does initially sample the lower energy regions, but as the simulation
progresses and those flatten out you move towards more uniform sampling.
The other thing I would add to these specific points is that a method like
metadynamics is easily expanded to an N-dimensional reaction coordinate
(although the amount of sampling required to explore this properly may make
it irrelevant in many cases) whereas I think this is more challenging for
many of the other methods. That being said, an additional point to what
Niklaus mentioned, is that while in theory you can just continue a
metadynamics or ABF run until it properly converges, if you've chosen
parameters you later realize where poor choices (e.g. bin widths for the
resolution of your PMF, too few samples for ABF, or too high of hills for
metadynamics), you will still most likely end up restarting the whole
thing. With umbrella sampling, however, you can always continue to
simulate and never worry that anything is a complete waste.
On the note of umbrella sampling and convergence, it's probably best
practice for each window to start from a set of coordinates you think are
different from what that window's equilibrium configuration will look like
and run two different simulations with differing random variables
(langevinThermostat for instance) and only use the portions of the
simulations that are identical (e.g. after the two different simulations
have converged upon one another) in your analysis. It means more
simulating, but it also means you can be more confident that you don't have
the wrong answer because of a slow degree of freedom creating a large
energy barrier as Niklaus mentioned.
On Mon, Mar 25, 2013 at 5:17 PM, Niklaus Johner <nij2003_at_med.cornell.edu>wrote:
> I have tried out several methods for several different problems over the
> last year, but my experience is still limited. I'll nevertheless give you
> my opinion. I'll won't consider the alchemical transformations here (FEP,
> thermodynamic integration,...)
> First of all I think SMD has a disadvantage over all the others. Namely
> if you don't pull slowly enough, you will largely overestimate the force
> and you won't have any way of increasing the sampling except restarting the
> simulation with a smaller pulling speed. With most other methods, if your
> free energy is not converged you can simply continue your simulation.
> Now in my experience, using SMD for non-equilibrium simulations and the
> Jarzynski type of equalities to calculate free energies doesn't work too
> well. I was calculating the binding free energy of a peptide to a membrane,
> but the friction of the water on the peptide would lead to large
> overestimates of the free energy. Better go for a very slow pulling force
> and very few repeats.
> Umbrella sampling seems to be a very safe bet in many cases. The main
> drawback is when using a CV that is highly degenerate (multiple states for
> a single value of the CV) and for which the energy barriers to interconvert
> from one state to the other at that particular value of the CV are large.
> Then you will basically be highly biased towards the starting structures
> that you use in each window. Think of the folding of a peptide, described
> by the radius of gyration for example, where it might be very difficult to
> interconvert between two distinct folded states without going through an
> unfolded state (large radius of gyration). In such cases methods like
> metadynamics or ABF, where you will cross the whole range of the CV many
> times will be more appropriate.
> For ABF and metadynamics, I think that a major difference is that in
> metadynamics you will mainly sample the low free energy regions, whereas in
> ABF you will get a more homogeneous sampling of the whole CV space.
> Metadynamics has the disadvantage of offering (to my knowledge) no real
> error estimate.
> Another method similar to metadynamics but where error estimates are
> straightforward, and allows for multiple totally independent simulations is
> "driven adiabatic free energy dynamics". This is nevertheless not available
> in namd but is included in the plumed plugin.
> Hope that helps,
> Niklaus Johner
> Weill Cornell Medical College
> Harel Weinstein Lab
> Department of Physiology and Biophysics
> 1300 York Avenue, Room D-501
> New York, NY 10065
> On Mar 25, 2013, at 3:56 PM, Thomas C. Bishop wrote:
> This is for a molecular dynamics class I'm doing. The question has come
> up "What's pro/con of diff methods and which works best when & where."
> Aron gave an excellent start.
> It would be great for experts to offer input and I"ll try to organize
> into something useful.
> On 03/25/2013 02:50 PM, JC Gumbart wrote:
> What are you trying to calculate specifically?
> On Mar 25, 2013, at 2:49 PM, Thomas C. Bishop wrote:
> Does anyone happen to have a "cheat-sheet" with the pros and cons of the
> diff free-energy methods or a review that provides this information.
> Thanks in advance,
> Thomas C. Bishop
> Tel: 318-257-5209
> Fax: 318-257-3823
-- Aron Broom M.Sc PhD Student Department of Chemistry University of Waterloo
This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:21:04 CST