rank_genes_groups logfoldchange different than seurat

[nagaBoulevard]

Hi!
I've noticed that the function rank_genes_groups calculate logFC differently than seurat.
https://github.com/theislab/scanpy/blob/5800db45dde0c2060ad0a9bed8ea931bd41da936/scanpy/tools/_rank_genes_groups.py#L207-L208

https://github.com/theislab/scanpy/blob/5800db45dde0c2060ad0a9bed8ea931bd41da936/scanpy/tools/_rank_genes_groups.py#L223

Thus the equation is
log(exp(mean(values))

while in Seurat is
https://github.com/satijalab/seurat/blob/96d07d80bc4b6513b93e9c10d8a9d57ae7016f9f/R/differential_expression.R#L175-L179

thus
log(mean(exp(values)))

I was thus wondering if this was intended, since it leads to different logFC values.

[Superoxide-Dismutase]

I found this problem too. Now logFC is still calculated in this way, that I am not satisfied with. When we are talking about average fold change of gene expression, the fold change of non-loged average expression is expected. In this way people get an intuitive feeling about how many times a gene is expressed compared with another group. So the expm() step must be done before the mean() step. Swap this order not only changes the logFC vaule, but also loses the biological meaning and doesn't make any sense.

[flying-sheep]

This is explained in #519 (comment)

In short: it is intentional, and is biased in the face of a lot of outliers, but much faster and otherwise very similar to a more accurate estimate.

If anyone wants to add an option to do it the other way, I’m open to it.

[eroell]

Hm, I see that this way has some advantages.

What I'd expect for the computation of logfoldchange is what @nagaBoulevard described, and I'd be opinionated to do it the other way (log(mean(exp(values))))

[flying-sheep]

I hope we can offer an option without making the code too complex. We need to benchmark if it’s possible to do the more accurate thing by default in scanpy 2.0.

I think doing it dynamically based on dataset size would be too unpredictable.

[ilan-gold]

@eroell does this only make sense to do for wilcoxon or also the t-test? I could see this making less sense there because the t test operates on means.