[gmx-users] free energy calculations using MBAR and gromacs

[Michael Shirts]

This looks like more of a pymbar question!

>  I am trying to calculate free energy of a system that involves
> disappearance of LJ particle at lambda=1 in explicit solvent. I ran the
> simulation at 20 different lambda points ranging from 0 to 1, using soft
> core potential. In order to use MBAR method (python implementation from
> Michael shirts ), I have to rerun the simulations, in order to get
> potential energies for each configuration at every lambda point.

That's correct.  Very soon (i.e., the code is in the CVS) it will be
possible to generate these energy differences during single runs, by
setting the keyword foreign-lambda.

> Now,
> here i get into trouble. If i evaluate the trajectories generated at
> lambda = 1 (particle completely decoupled) using the potential function
> at lambda = 0, i get very large energy values,

Yes; if you look at the example in pymbar, there are lots of values of
E0 - E1 of around 10^14.

> which quite likely are
> responsible for MBAR's failure to converge.

That shouldn't be the case.  These energy values turn into weights,
which are essentially zero, and thus don't affect MBAR.  This sounds
like is more of a pymbar question that a Gromacs question!

> The program works well if i
> calculate free energy to/from an intermediate. ie lambda 0 -> 0.5 and
> 0.5->1.0

Hmm. It should work well in either case!

> As the potential energies for the configurations generated at lambda=1,
> evaluated using potential energy function at lambda =0, is expected to
> be infinite (due to Van der Waals overlaps between solvent and LJ
> particle),

Very large but finite, but yes.

> Is it even possible to apply MBAR method to such cases,
> without splitting the analysis to an intermediate lambda value? Did
> anyone ran into similar problems, or am i missing something?

Yes, it is possible.  If you send a copy of 1) the data and 2) the
scripts you use to interface with MBAR, I can take a look to try to
see what the matter will be.

Best,
Michael