msmbuilder-legacy Update tICA docs

So there's a file docs/tICA/tICA.pdf that shares a lot in common with http://msmbuilder.s3-website-us-east-1.amazonaws.com/theory/tICA.html

However, there are several key differences:

HTML version is missing the actual MSMBuilder commands

I also noticed that the PDF version instructs users to download a special fork and branch, which I believe is no longer necessary, right?

Feb 05 '14 16:02 kyleabeauchamp

OK I guess what happened was that Robert split the tICA doc into theory and application but never ported the application side of things.

I wonder if it makes more sense to merge them. Not sure.

Feb 05 '14 16:02 kyleabeauchamp

Do you have a preferred way to generate the atom pairs? Otherwise, I vote that we include a tiny script inside the tutorial (probably at the very end, in a FAQ):

import itertools
import numpy as np
import mdtraj as md

trj = md.load("./system.subset.pdb")

top, bonds = trj.top.to_dataframe()
atom_indices = np.where((top.name == "CA") & (top.resSeq >= 139) & (top.resSeq <= 175))[0]
atom_pairs = list(itertools.combinations(atom_indices, 2))

np.savetxt("./AtomPairs.dat", atom_pairs, "%d")

Feb 05 '14 16:02 kyleabeauchamp

Also, I wonder if it might make sense to set stride = 1 for the "beginner" TICA tutorial. Otherwise there's just a lot of parameters exposed to users.

I know you've found the 1 / 10 stride optimal, but I wonder if it make sense to ignore that fact for illustrative purposes.

Feb 05 '14 16:02 kyleabeauchamp

Also: the "Drawbacks of TICA" section actually applies to all MSM-like forms of dimensionality reduction, so it might not be necessary.

Feb 05 '14 17:02 kyleabeauchamp

thanks for checking this out, I'm going to make these changes (hopefully later today)

Feb 05 '14 17:02 schwancr

Also: IMHO eliminate the ProjectInfo.yaml inputs to scripts, as they get set by default. I think that will help people stay focused on the tICA-specific details.

Feb 05 '14 17:02 kyleabeauchamp

If you do end up merging the theory + applications into a single sphinx file, could you move the link to sit under the "Documentation" section on the main docs page?

Feb 05 '14 18:02 kyleabeauchamp

Again, let me know if you run out of time and I can file a PR for some of this stuff.

Feb 05 '14 18:02 kyleabeauchamp

The reason I split them was because the original tICA tutorial latex file I was working from had all this info about downloading a different branch, and I wasn't sure which parts were or were not relevant currently. But the theory I knew was current.

I don't think having the theory and practice pages separated is a bad idea, especially if we can put in links between them.

Feb 05 '14 19:02 rmcgibbo

So should I use k-centers or k-medoids when clustering my tICA results? Because we're working in the eigenvector space, I imagine that either one of the following is true:

k-centers neglects equilibrium density because we're working in right eigenvector space
k-medoids double-counts equilibrium density because we're working in left eigenvector space

Feb 06 '14 18:02 kyleabeauchamp

In practice, I've found that the hybrid k-medoids that msmb implements doesn't change things drastically. If you wanted to do k-means, however, you could probably gain a lot.

I don't know the rigorously right way to do it, but for instance, when I used Ward clustering, I could build a 20 state model (just from clustering) that gave me the same model (with slightly faster timescales) as building a 1,000 state model with k-centers.

On Thu, Feb 6, 2014 at 10:59 AM, kyleabeauchamp [email protected]:

So should I use k-centers or k-medoids when clustering my tICA results? Because we're working in the eigenvector space, I imagine that either one of the following is true:

k-centers neglects equilibrium density because we're working in right eigenvector space

k-medoids double-counts equilibrium density because we're working in left eigenvector space

Reply to this email directly or view it on GitHubhttps://github.com/SimTk/msmbuilder/issues/324#issuecomment-34357495 .

Feb 06 '14 19:02 schwancr

Thanks. Have you ever done k-means? E.g. do I have to update Cluster.py?