- Contemporary messages sorted: [ by date ] [ by thread ] [ by subject ] [ by author ] [ by messages with attachments ]

From: Jason Swails <jason.swails.gmail.com>

Date: Mon, 18 Jan 2016 16:12:16 -0500

On Mon, Jan 18, 2016 at 3:03 PM, Melisa Averina <maverina.ucsd.edu> wrote:

*> Hello,
*

*>
*

*> I am trying to simulate a very large protein but I am only interested in
*

*> the dynamics of only a small region of the protein that binds to an
*

*> external molecule. I was wondering if we can save an enormous amount of
*

*> computational effort by holding the entire protein completely fixed to its
*

*> original structure except for the small portion I am interested in, i.e.,
*

*> integrate the equations of motion of only those atoms belonging to this
*

*> small portion, so most of the inter-atomic forces are not even evaluated
*

*> (saving a lot of time).
*

*>
*

*> I have read about belly, freeze, and shake but I believe they use some
*

*> kinds of restraints, as so it won't help in accelerating the simulation
*

*> itself.
*

*>
*

You need to define what you mean by "accelerate the simulation". If you

constrain a large fraction of the degrees of freedom (which is what belly

does -- it uses constraints, not restraints), then you substantially reduce

the size of phase space. This reduces the amount of sampling needed to

characterize the relevant parts of phase space, meaning you can more fully

characterize your model with fewer time steps.

Indeed, this results in a substantially cheaper model. It also changes

that model a lot and makes it a lot less realistic. So the added

simplicity may (and probably will) come at the cost of accuracy.

However, you seem to be equating "accelerating the simulation" with taking

less time per time step. This is impossible in Amber. There's a bigger

problem, though, with the underlying assumption that will significantly

reduce the feasibility of this approach -- the commonly-used potential

energy functions in molecular dynamics are not pairwise decomposable, and

so you cannot simply omit certain pairs of interactions to save time in the

general case.

For GB or PB, the GB radii or dielectric boundary affects all pair-pair

interactions, but are determined by all atoms in the system. So you can't

eliminate redundant interactions as much as you think you can. For PME,

the reciprocal space calculation itself is entirely non-local, as it is

performed in Fourier space. Sure, you can simplify some of the

direct-space calculation and eliminate some interactions, but you'll still

be limited by the PME part.

The savings in your proposed scheme would come not from the reduced cost of

each time step, but rather the vast reduction in the number of degrees of

freedom. But they would come with a cost.

HTH,

Jason

Date: Mon, 18 Jan 2016 16:12:16 -0500

On Mon, Jan 18, 2016 at 3:03 PM, Melisa Averina <maverina.ucsd.edu> wrote:

You need to define what you mean by "accelerate the simulation". If you

constrain a large fraction of the degrees of freedom (which is what belly

does -- it uses constraints, not restraints), then you substantially reduce

the size of phase space. This reduces the amount of sampling needed to

characterize the relevant parts of phase space, meaning you can more fully

characterize your model with fewer time steps.

Indeed, this results in a substantially cheaper model. It also changes

that model a lot and makes it a lot less realistic. So the added

simplicity may (and probably will) come at the cost of accuracy.

However, you seem to be equating "accelerating the simulation" with taking

less time per time step. This is impossible in Amber. There's a bigger

problem, though, with the underlying assumption that will significantly

reduce the feasibility of this approach -- the commonly-used potential

energy functions in molecular dynamics are not pairwise decomposable, and

so you cannot simply omit certain pairs of interactions to save time in the

general case.

For GB or PB, the GB radii or dielectric boundary affects all pair-pair

interactions, but are determined by all atoms in the system. So you can't

eliminate redundant interactions as much as you think you can. For PME,

the reciprocal space calculation itself is entirely non-local, as it is

performed in Fourier space. Sure, you can simplify some of the

direct-space calculation and eliminate some interactions, but you'll still

be limited by the PME part.

The savings in your proposed scheme would come not from the reduced cost of

each time step, but rather the vast reduction in the number of degrees of

freedom. But they would come with a cost.

HTH,

Jason

-- Jason M. Swails BioMaPS, Rutgers University Postdoctoral Researcher _______________________________________________ AMBER mailing list AMBER.ambermd.org http://lists.ambermd.org/mailman/listinfo/amberReceived on Mon Jan 18 2016 - 13:30:03 PST

Custom Search