Native python implementation for NB-GLM instead of edgeR

[emdann]

Gathering some suggestions:

• Patsy for model matrix encoding https://patsy.readthedocs.io/en/latest/overview.html
• GLM implementation in pymc3 https://docs.pymc.io/notebooks/GLM-poisson-regression.html

[emdann]

statsmodels https://www.statsmodels.org/stable/generated/statsmodels.formula.api.negativebinomial.html#statsmodels.formula.api.negativebinomial

[MikeDMorgan]

tensorflow might be useful for GLMs: https://www.tensorflow.org/probability/examples/Generalized_Linear_Models
I've not tested/used this, but it gives a stable API.

[emdann]

batchGLM / diffxpy from theislab

[Zethson]

Theislab dude here. Just wanted to say that we have some people working on batchGLM again and I think that it would be a good choice for you in the future.

[emdann]

Hi @Zethson, thanks for the heads up. I actually have some bandwidth now to look into this and try a few options out.

[Zethson]

Hi @Zethson, thanks for the heads up. I actually have some bandwidth now to look into this and try a few options out.

Well, batchglm will still take a few months to be usable and well documented, but feel free to take a look nevertheless :) (development is happening on a couple of branches so don't look at master).

[picciama]

Hi @Zethson, thanks for the heads up. I actually have some bandwidth now to look into this and try a few options out.

Well, batchglm will still take a few months to be usable and well documented, but feel free to take a look nevertheless :) (development is happening on a couple of branches so don't look at master).

batchglm developer here with a minor update on this for @emdann and @Zethson: The python port of edgeR using batchglm as backend is implemented for your purposes but I didn't have the time for testing / deployment yet. It's probably going to take till January until I find time for that but let me just say that for my small testing dataset it is already pretty much exactly reproducing edgeR's results.

[Zethson]

Excellent @picciama ! Thank you very much! Really think that a Python native edgeR will be super useful for the community. Just a big bonus that we can add this to Milopy.

CC @ivirshup

[Zethson]

BTW @picciama this is where the new batchGLM based implementation should then end up in -> https://github.com/theislab/pertpy/blob/development/pertpy/tools/_milopy.py#L306
(Just another if case)