Re: [AMBER] clustering

From: Sowmya Indrakumar <>
Date: Thu, 2 Feb 2017 08:46:57 +0000

Dear Ravi,
Thank you for this email. It was very informative.
I now understand, avergelinkage is a heirarchial agglomerative clustering method using RMSD metrics. (bottoms up aprroach) and hence there is no consideration of any reference initial structure.
Thank you once again.
From: Ravi Abrol []
Sent: Thursday, February 02, 2017 12:08 AM
To: AMBER Mailing List
Subject: Re: [AMBER] clustering

Hello Sowmya,

You are asking a couple of different questions.

First, how the clustering method partitions the trajectory frames into
different clusters and how the
different clusters are ordered during output. This depends on the method
used and in your case,
the cpptraj documentation will have that answer.

Second is the use of an "average" structure. If large conformational
changes are going on, then there
is no good average structure. If you really need an average structure, the
"raw" average structure
obtained by averaging x,y,z over all trajectory frames is almost always
unphysical as Bill Ross pointed out.
A "realistic" average structure can be obtained by comparing all the
trajectory frames to the "raw"
average structure and picking "the" frame that has the smallest RMSD to the
"raw" average structure.


On Wed, Feb 1, 2017 at 3:41 AM, Sowmya Indrakumar <>
> Dear Daniel and Bill,
> Thank you for your suggestions.
> I was expecting RMSD to increase with the cluster number, for instance
> RMSD of the representative structure to the average structure should be the
> least considered to the last cluster representative structure to the
> average). Since here, the frames are partitioned into distinct clusters, we
> don't take into account any reference structure.
> But as you suggested, comparing to the average structure doesn't make
> sense if there is a lot of changes.
> It is still not clear how frames are put into distinct cluster, does it
> compare all the frames and then assign them to various clusters?
> Thanks
> Regards
> SOwmya
> ________________________________________
> From: Bill Ross []
> Sent: Tuesday, January 31, 2017 7:11 PM
> To: AMBER Mailing List
> Subject: Re: [AMBER] clustering
> Not sure if it has been pointed out, but average structures from
> trajectories can be quite unrealistic if the system has interesting
> changes going on.
> Bill
> On 1/31/17 9:59 AM, Daniel Roe wrote:
> > On Tue, Jan 31, 2017 at 10:03 AM, Sowmya Indrakumar <>
> wrote:
> >> running the above commands gives me 5 representative structures. It is
> not very clear to me what is it taking as the reference structure.
> Comparing the five generated representative pdb structure to the average
> pdb (from the trajectory) does not give show any trend in RMSD.
> > The representative structure of each cluster is the frame with the
> > lowest cumulative "distance" (in your case RMSD) to every other frame
> > in the cluster. A cluster representative is just that; it's supposed
> > to be representative of what that cluster looks like (although this
> > may not always be the case). I can see how comparing representative
> > structures to the average structure of the cluster could be
> > informative ('avgout', 'avgfmt'), but it's not clear to me what
> > insight you can gain from comparing representatives to the overall
> > average structure.
> >
> > -Dan
> >
> _______________________________________________
> AMBER mailing list
> _______________________________________________
> AMBER mailing list
AMBER mailing list
AMBER mailing list
Received on Thu Feb 02 2017 - 01:00:03 PST
Custom Search