Two major sources of error characterize a Markov state model. The first is
the convergence of the dataset -- we can only model the processes that we've
simulated, in some form or another. When the sampling is insufficient, its
not like the MSM can make something out of nothing. This sampling error is
hard to model, because we *don't know* whats out there. The best that we can
do is assess how "densely" sampled the data we've seen is. Do we have
transitions that we've only seen once? If we split the dataset in half, are the
halves look consistent with one another?

The second major source of uncertainty is uncertainty in the parameters. These are the number of clusters, their locations and size, and transition probabilities between them (also the lag time, but let's leave that aside for now).

To really assess these errors, we need an approach that models them all explicitly. This is challenging though -- our traditional clustering approaches are parametric in the number of states, so they don't allow us to naturally express our uncertainty there. We're going to need to go nonparametric.

Here's the idea: Small peptide (ala2 or ala5). Dirichlet process gaussian mixture model "clustering" (nonparametric in number of states), with a Gibbs sampler so that we can sample from the posterior over clusterings. For each clustering, sample transition matrices with the MCMC engine that Kyle's been working on. We would have to do the whole thing in probably the projected backbone dihedral space. It would be nice to use the von Mises distribution instead of the gaussian to avoid going into the sin/cos space doubling the number of variable.

I think going nonparameteric in the number of states is pretty key. This is probably going to be really expensive, and tuning the settings on the samplers could definitely be a nightmare (now there's a whole separate convergence issue to worry about), but it would be nice do a careful accounting for the uncertainties. Other approaches that you might do don't really model the uncertainty in the clustering parameters.

A simpler approach might be to run a bootstrap on the actual trajectory data. That's a good option too, but not as elegant. You don't get uncertainty in the number of states, and there are so many issues with the bootstrap on timeseries.