Re: ABF non-uniform sampling

From: Jérôme Hénin (jhenin_at_ifr88.cnrs-mrs.fr)
Date: Fri Apr 16 2010 - 04:48:17 CDT

Hi Patrick,

Having sampling vary from 25000 to 40000 samples per bin is not so
bad, it still means that the whole range is well-sampled. Completely
uniform sampling is unlikely, simply because of the stochastic nature
of the simulation.

Instead, the discrepancy probably comes from a different problem: the
convention used to define the PMF, including or excluding the
geometric entropy term (Jacobian correction). The Tcl routines used to
make a special case for the distance coordinate, so that PMFs would
match the usual convention. In colvars, we aimed for a more consistent
and systematic behavior, so by default all PMFs are output in the
"probabilistic convention", including the geometric term.

The colvars will follow this convention if you add the "hideJacobian"
keyword (see the user's guide for a full description). In the case of
a distance r, this term is simply equal to -2*RT*ln(r), so you can
easily be able to recover the corrected PMF by subtracting this
quantity.

The current state of the ABF tutorial does not mention all this,
though, so thank you for reminding me. It will be added to the next
revision of the tutorial.

Best,
Jerome

On 16 April 2010 05:19, Patrick Yee <pyee_at_nd.edu> wrote:
> Hi NAMD users,
>
> I am having a problem with non-uniform sampling using ABF and colvars.
> I did not have these sampling problems when using the TCL scripts for
> ABF in NAMD 2.6. My sampling increases fairly linearly until the
> simulation ends and even long simulations (10-20 ns) did not have
> uniform sampling. Does anyone have any idea of why the sampling is not
> uniform? I've given more details below.
>
> I've just started using colvars (instead of the NAMD 2.6 TCL scripts)
> for calculating a PMF with the ABF method. To test if the PMF would
> work on my systems when I used the TCL script ABF method, I first
> tried finding the PMF on pair of methane molecules in water and then
> on Na+ and Cl- ions in water (the latter being from the tutorial); the
> results matched very closely what I'd found from other MD simulations.
> When I tried these same simulations in NAMD 2.7b2 with colvars, I
> didn't get the correct PMF as I had using ABF with TCL scripts, most
> likely because the sampling is not becoming uniform as the simulation
> progresses. Note that with colvars I used fullSamples = 500, which was
> the same as with the TCL script. I also thought the simulation might
> not have been allowed to proceed long enough, but even after 10
> million steps at 1fs/step, the sampling was very non-uniform. Sampling
> ranges from 40000 when the groups are close and moves fairly linearly
> to about 25000 as the groups move apart. I attached the colvars input
> file I am using below. Again, please tell me why the sampling is not
> becoming uniform.
>
> Thank you in advance for your help.
>
> Colvarstrajfrequency        2000
>
> colvar {
>
>    name              IonDistance
>    width             0.1
>    lowerboundary     0.05
>    upperboundary     13.00
>    lowerwallconstant 10
>    upperwallconstant 10
>
>    distance {
>        group1 {
>        atomnumbers    { 1 }
>        }
>        group2 {
>        atomnumbers { 2 }
>        }
>    }
> }
>
> abf {
>    colvars           IonDistance
>    fullSamples       500
> }
>
>

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:55:41 CST