Integration with msmtools

[mpharrigan]

To facilitate greater code re-use, msmbuilder will start using msmtools (https://github.com/markovmodel/msmtools) for algorithms and functionality. msmbuilder will act as a "front end" to deal with UI, dataset management, documentation, and others.

PyEMMA (https://github.com/markovmodel/PyEMMA) has already been refactored in this model. PyEmma and MSMBuilder will both be built on mdtraj (for trajectory processing, managed by pandegroup) and msmtools (for msm algorithms, managed by prof. Noe's group)

Let's start thinking about how to transition some algorithms to msmtools. Something like the following:

1. Identify algorithms in msmtools that are not present in msmbuilder and expose in msmbuilder
2. Identify algorithms that are better(?) in msmtools and switch our implementation to that
3. Identify algorithms in msmbuilder that aren't in msmtools, contribute them, and switch over (yay collaboration!)
4. Switch over remaining algorithms

cc @rmcgibbo @franknoe @marscher

[franknoe]

I can only comment on point 3:

• might include the rate matrix estimation or parts of it (and that discussion has already started between Fabian and Robert).
• you have actually implemented Hao's reversible transition matrix estimation algorithm from Prinz et al JCP 2011, which we have never ported to python/C so far. That might be useful to contribute. Have you guys ever tested whether that is consistently faster than the fixed point iteration? We haven't.
• HMM stuff might be very interesting, but that is off-topic as it would rather go to bhmm and requires a few refactoring events to occur first
• there's probably more that I am not aware of.

Am 17/09/15 um 21:26 schrieb Matthew Harrigan:

To facilitate greater code re-use, msmbuilder will start using msmtools (https://github.com/markovmodel/msmtools) for algorithms and functionality. msmbuilder will act as a "front end" to deal with UI, dataset management, documentation, and others.

PyEMMA (https://github.com/markovmodel/PyEMMA) has already been refactored in this model. PyEmma and MSMBuilder will both be built on mdtraj (for trajectory processing, managed by pandegroup) and msmtools (for msm algorithms, managed by prof. Noe's group)

Let's start thinking about how to transition some algorithms to msmtools. Something like the following:

1.

Identify algorithms in msmtools that are not present in msmbuilder
and expose in msmbuilder

2.

Identify algorithms that are better(?) in msmtools and switch our
implementation to that

3.

Identify algorithms in msmbuilder that aren't in msmtools,
contribute them, and switch over (yay collaboration!)

4.

Switch over remaining algorithms

cc @rmcgibbo https://github.com/rmcgibbo @franknoe https://github.com/franknoe @marscher https://github.com/marscher

— Reply to this email directly or view it on GitHub https://github.com/msmbuilder/msmbuilder/issues/653.

---

Prof. Dr. Frank Noe Head of Computational Molecular Biology group Freie Universitaet Berlin

Phone: (+49) (0)30 838 75354 Web: research.franknoe.de

Mail: Arnimallee 6, 14195 Berlin, Germany

[rmcgibbo]

@franknoe the msmbuilder transmat estimator looks a little faster to me, but not hugely different: https://gist.github.com/rmcgibbo/85ea5f0bde9d6b79b9dc

[franknoe]

Thanks for the quick check! I think it's worth contributing it. I suggest to allow switching between methods on the API level (e.g. msmtools.estimation.transition_matrix) using a "method='default'" flag.

Since I'll be traveling the next two days, it would be best to coordinate this with Benjamin and/or Fabian. They have built the estimation algorithms in msmtools.

Am 18/09/15 um 00:34 schrieb Robert T. McGibbon:

@franknoe https://github.com/franknoe It looks a little faster to me, but not hugely different: https://gist.github.com/rmcgibbo/85ea5f0bde9d6b79b9dc

— Reply to this email directly or view it on GitHub https://github.com/msmbuilder/msmbuilder/issues/653#issuecomment-141256287.

---

Prof. Dr. Frank Noe Head of Computational Molecular Biology group Freie Universitaet Berlin

Phone: (+49) (0)30 838 75354 Web: research.franknoe.de

Mail: Arnimallee 6, 14195 Berlin, Germany