Types of emissions for the Hidden Markov Model

[igordot]

MACS3 3.0.2 introduced the option --hmm-type for 'gaussian' or 'poisson' HMM (#635).

The documentation just says "the Gaussian emission can be described by mean and variance for each state, while the simpler Poisson only needs the lambda value".

In one of the issues (#670), there is a mention that "the new v3.0.2 version ... provides a nicer poisson emission model version of HMM". In another (#636), "the Poisson Emission model ... (compared with the Gaussian Emission model) generated results that were similar to those of the Java implementation.

One might assume that the Poisson model is more reliable given that it's similar to the original Java implementation that has been around for much longer and more extensively tested.

Is it possible to get some clarification about the tradeoffs of the two versions and how to select between them?

[taoliu]

We can't implement exactly the same thing as the Java version HMRRATAC in Python.o Our goal is to implement it using Python libraries instead. In Python, we choose to use hmmlearn library and aim to provide options as many as possible for users to tweak the behavior of macs3 hmmratac. So far, the simple Poisson emission model seems to generate the results close to the java version on the same small testing data, although there is no such emission model for the Java version. We will try to do a thorough evaluation using the same data in HMMRATAC paper and post it on MACS3 website later.