Speaking of clustering, I am getting interesting 3-mode 'natural'
clustering results using Poincare spherical space for a distance
function using photo histograms:
http://phobrain.com/pr/home/poincaredesc.html
Bill
On 9/25/17 2:47 AM, Thomas Cheatham wrote:
>> I am doing cluster analysis using the following script from:
>> http://www.amber.utah.edu/AMBER-workshop/London-2015/Cluster/
> One problem with following a workshop script is that you do not have the
> instructors in the audience to help out and often such scripts are
> explanatory for a particular data set and not necessarily generally
> useful. Note that there are many clustering approaches, each with
> different assumptions.
>
> The first thing I would do is create a 2D-RMSD plot to see if there are
> clear clusters and to show me what to expect. If there are no blocks
> visible, likely the RMSD vs. time is just increasing and you are not
> re-visiting conformations.
>
> Try to understand what the algorithms are doing...
>
> In your case, you have sieve 100 so only 200 frames are looked at the for
> the clustering. You have minpoints of 25; since no clusters are being
> formed, in those 200 frames 25 of them (1/8th) are not falling into a
> single cluster (hence no cluster is formed). Finally, the time to do the
> clustering was only 10 s (compared to much longer for reading trajectory)
> and memory usage was low, so why not cluster more frames? My guess is
> that your RMSD is increasing over time... I would try sieve 10.
>
> Also, other clustering algorithms do not have minimum sizes; try
> hierarchical or k-means or some other algorithm too. Clustering is kind
> of an art and you need to experiment a bit...
>
> --tec3
>
>> Minimum pts to form cluster= 25
>> Cluster distance criterion= 0.900
> (maybe set cluster distance criterion higher?)
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Sep 25 2017 - 03:30:02 PDT
