Re: [AMBER] PCA trajectories give unrealistic structures

From: Daniel Roe <>
Date: Wed, 9 Nov 2016 15:33:36 -0500


On Wed, Nov 9, 2016 at 11:06 AM, Gulsevin,Alican <> wrote:
> When I run PCA with two structures, what I get is a nice output trajectory where I see loops moving around.

Some more details here would help. When you say "two structures", do
you mean that you are performing PCA on the combination of two
separate trajectories? If you provide your exact cpptraj input (and
ideally output) that would give us more insight into what you are
trying to do and what is actually happening.

> However, when I use five structures, what I see in the trajectory is an unrealistically contracted protein growing into a normal size protein, which clearly is not a representation of what's really going on. I have two questions at this point:

I'm assuming here you are referring to the pseudo-trajectories
obtained with the 'modes trajout' command. I would like to point out
this paragraph in the tutorial:

"One important thing to keep in mind is that while PCA is useful for
gaining insight into the dynamics of a system, the actual motion of
the system throughout the course of a simulation is almost always a
combination of the individual PCs. So while motion along a single PC
might show a transition, it is usually not actually how the system
undergoes that transition."

In other words, motion observed along a single PC does not necessarily
reflect the actual motions undergone by the system. Principal
components are vectors which explain variances in data - in the case
of PCs calculated from a coordinate covariance matrix these are
collective motions of the system. It's useful to show you in
particular what low-frequency (corresponding to high eigenvalue)
motions are being undergone by your system, but the actual motions are
only fully described by various combinations of *all* the principal

> 1) Why do I keep seeing this weird trajectory with 5 ligands, but not with 2? All structures are referenced to the same pdb structure and prmtop files are the same (because proteins are identical in all cases). Is CPPTRAJ PCA designed to compare two structures only by its nature, or is there another step necessary in addition to what's described because I'm using multiple MD results?

This will be easier to answer once we see your cpptraj input/output.

> 2) Is there a way to get a pretty PC1 vs PC2 kind of plot using CPPTRAJ? A reply to a past question suggests keyword "projection", but I wonder if it's compatible with the kind of analysis I described, again for a larger number of trajectories.

Do you mean you want to plot the 2D histogram of the projection of
your trajectories along PCs 1 and 2? If yes, then you will want to use
the 'projection' command to generate the projection data sets, and
then 'hist' to plot those. See supporting information here

for examples of using the 'projection' command in conjunction with
'hist'. Also see the cpptraj section of the Amber 16 manual for full
details of the commands and examples.


PS - If you haven't already, I recommend you read some of the
literature that uses PCA on MD data (an early one that comes to mind
is but there is no
shortage of others).

Daniel R. Roe
Laboratory of Computational Biology
National Institutes of Health, NHLBI
5635 Fishers Ln, Rm T900
Rockville MD, 20852
AMBER mailing list
Received on Wed Nov 09 2016 - 13:00:02 PST
Custom Search