The order of the clusters in the log file obtained from GROMACS clustering

User discussions
1

GROMACS version: 2020.4
GROMACS modification: No
Dear GROMACS user,
I have a quick question about the output log file obtained from GROMACS clustering. Specifically, this is the command that I was using (i was set as 9 before the command execution):

mpirun -np 1 gmx_mpi cluster -f ../MD/*.xtc -s ../MD/*tpr -n *ndx -cl glycoform_${i}_ACS_cluster_dt250.pdb -g cluster.log -cutoff 0.10 -dt 250

After the command is finished, in the file cluster.log, I found the beginning part of the file look like this:

Using linkage method for clustering
Using RMSD cutoff 0.1 nm
The RMSD ranges from 0.0421666 to 0.734086 nm
Average RMSD is 0.263982
Number of structures for matrix 8001
Energy of the matrix is 85.5332.

Found 540 clusters

Writing middle structure for each cluster to glycoform_9_ACS_cluster_dt250.pdb

cl. | #st  rmsd | middle rmsd | cluster members
  1 |   1       |      0      |      0
  2 |   1       |    250      |    250
  3 |   2  0.097 |    500 .097 |    500    750
  4 | 3545  0.179 | 386000 .142 |   1000   3250   3500   3750   4000   4250   4750
    |           |             |   5250   5500   5750   6500   6750   7000   7250
    |           |             |   7500   7750   8000   8250   8500   8750   9250
    |           |             |   9500   9750  10000  10250  10500  11000  11250
    |           |             |  11500  11750  12000  12250  12500  12750  13000
    |           |             |  13250  13500  13750  14000  14250  14500  14750

I was confused since I thought that the clusters in the log file should be ordered depending on their sizes, from the biggest cluster to the smallest clusters. However, as shown above, the first 3 clusters had 1, 1, and 2 members, while the fourth had 3545 members. I’m wondering if my interpretation is wrong and if the most representative structure would be the medoid of cluster 4 instead of cluster 1. Thank you!

2

Dear GROMACS,
Also, I’m wondering how I can find the optimal cutoff distance for clustering. My current understanding is that the cutoff distance should be small enough such that there is a reasonable number of clusters (<40) and there should be an even smaller of clusters that contain around 90% of the frames. I’m curious if there is any other rule of thumb that should be applied here as well. Thank you!

Best,
Wei-Tse