Re: [AMBER] Principal Component Analysis

From: Jason Swails <jason.swails.gmail.com>
Date: Wed, 8 Jun 2016 22:36:05 -0400

On Wed, Jun 8, 2016 at 9:10 PM, Kat G <katwin86.gmail.com> wrote:

> Hi all,
>
> I am trying to use PCA to study simulations of one protein in its apo form
> and holo form with 3 different ligands. After doing PCA separately for each
> system, the correlation is quite different among systems, compared with
> what we are expecting from their binding free energy. For example, the
> strongest binding activator would mostly stabilize the protein, and then
> having the least number of eigenvectors to represent 90% of all possible
> modes.
>
> I suppose that it may be wrong when I made a comparison based on different
> types of eigenvectors in 4 systems. I am not sure that these eigenvectors
> reflect the same mode of action in 4 cases or not. How could I get the same
> type of eigenvectors in both apo and holo-protein to make comparison
> possibly useful.
>
> I have heard about combined-PCA by concatenating different trajectories.
> Could it be done by cpptraj.
>

​Sure, just trajin all the trajectories you want to include in the PCA
calculations.

An alternative thing you can do is to do PCA on *one* trajectory (just of
the complex), and then use those eigenvectors as the "principal components"
when plotting the projections of the other trajectories. That's often
what's done to compare two independent simulations of the same system to
check that they give the same converged answer.

If the ligands are small, the vectors with the largest eigenvalues for one
trajectory are likely to represent a large fraction of the covariance for
the others as well, even if they are not the largest contributors.

You can always look at the similarity between the PCA eigenvectors of the
top couple modes for each trajectory by computing the inner product of them
and seeing how large that inner product is (relative to the magnitude of
the smaller of the two eigenvectors).

HTH,
Jason

-- 
Jason M. Swails
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jun 08 2016 - 20:00:02 PDT
Custom Search