Speed up Milo tutorial

[Zethson]

Description of feature

Currently takes 30mins+ to run which is too long. Maybe subsample or something else. It took a hit when we switched to pydeseq2 but I don't want to go back.

[FrenkT]

I had the same issue when I switched between two apparently similar datasets. I did some investigations and I think the reason is this behaviour of fit_size_factors:

 Uses the median-of-ratios method: see :func:`pydeseq2.preprocessing.deseq2_norm`, unless each gene has at least one sample with zero read counts, in which case it switches to the ``iterative`` method.

It's the iterative method that is very slow (which makes sense looking at the code). Only solution seems to be subsampling, or potentially a larger nhood size, so that it increases the probability that at least one nhood covers all the samples.

[Zethson]

Ahh, thank you! Would you be interested in playing around https://github.com/scverse/pertpy-tutorials/blob/main/milo.ipynb and see whether you can make it use the non-iterative method?

[FrenkT]

Hi @Zethson , I'll give it a go and will report back.