Re: [AMBER] Trajectory determination and directionality in TI with softcore potentials

From: Ying-Chieh Sun <>
Date: Wed, 30 Oct 2013 16:43:06 +0800

My quick comment to your question is ... the 'truncation' in numbers may
give deviation you mentioned. Therefore, there is tolerance in difference of
calculated <dV/d lambda> statistically. Ying-chieh Sun

-----Original Message-----
From: Ryan Muraglia []
Sent: Wednesday, September 18, 2013 3:39 AM
Subject: [AMBER] Trajectory determination and directionality in TI with
softcore potentials


I am using thermodynamic integration in AMBER 11 with softcore potentials,
and I am finding curious behavior that I cannot explain by reading the
manual or tutorials, so I was hoping that this would be an appropriate place
to seek an explanation or a suggestion for additional resources to consult.

I have been trying to compute the free energy difference of the mutation of
an alanine residue (in an alanine tripeptide, AAA) to a valine residue
(making the tripeptide AVA) in a box of TIP3P water as a test case. I can
share my input files if they will help. The masked atoms are the sidechains
of the residues that are being swapped.

One of my sanity checks was to run the TI scheme "backwards," meaning that
in the groupfile, I flipped the order of the sander calls. I was expecting
that the dv/dlambda curves for the forward and backward processes would
mirror each other (with one taking on negative values). For example I would
expect that AAA -> AVA for lambda=0.25 would take on the negated value for
average dv/dl as AVA -> AAA for lambda=0.75, and that the values at
lambda=0.5 would be exact opposites.

I found that my expectation was not met. I used linear mixing of the
potentials (klambda=1) and I used the same, default, fixed random seed for
the Langevin dynamics (ig=71227) for each simulation, so I am not sure where
the deviation from my expected behavior is coming from. As I understand it,
0.25 "forward" should be evolving according to the same mixed potential as
0.75 "backward." Digging deeper, I found that the trajectories of the two
processes did not match, which could cause the discrepancy in dv/dlambda
values down the line.

All of this background is simply to ask, how are the trajectories determined
in TI? There is obviously something more nuanced than simply stating that
the system evolves according to a mixed potential. Do non-masked atoms share
the same trajectory in the two endpoint states (V0 and V1)? How is
information from the two processes representing V0 and V1 combined to
determine atomic positions in the next step? And most importantly, is this
discrepancy to be expected, and if so, can it be characterized in a
meaningful manner, or is it a byproduct of the implementation?

Thank you very much for reading. Any insight or suggestions for further
reading would be greatly appreciated.

Ryan Muraglia
AMBER mailing list
AMBER mailing list
Received on Wed Oct 30 2013 - 02:00:09 PDT
Custom Search