Cluster inputs for calculating size factors?

[Demond-dev]

Hi,

In the R scran workflow it is suggested to perform rough clustering of data and provide that as input so that size factors are calculated in each cluster as opposed to size factors being calculated across all cells. Is there an equivalent to that in scranpy? Would the closest comparison be to provide the clusters as the block argument in the center_size_factors function? Or would the best approach be to perform a rough clustering before hand and run scranpy on the objects subset by cluster?

Thanks in advance for any clarification!

[LTLA]

scranpy doesn't actually implement the scran deconvolution algorithm. The normalization approach in scranpy.analyze() just uses the centered library sizes as the factors, so there's no need to do any clustering.

With the benefit of hindsight, I would say that the deconvolution approach isn't worth the effort for the (vast?) majority of analyses:

• Composition biases do not really affect the ability of downstream analyses to distinguish between cell populations. If you ever have significant composition bias, it means you have even bigger DE between your subpopulations, so there's no need to correct the bias in order for the subpopulations to be distinguished.
• A "big" composition bias would be 30-50%, i.e., ~1.5-2-fold. Typical markers would change by at least 8-10-fold between clusters. The bias won't affect your ability to identify relevant markers, and even if they do reduce the magnitude of the log-fold changes, the ranking of the markers shouldn't change.
• The deconvolution method relies on a semi-decent clustering to obtain accurate size factors. But if your clustering is already decent, you might as well just stop there rather than doing another round of your analysis with different size factors.
• If you really do care about the accuracy of the log-fold changes between populations, it would be much more accurate to compute pseudo-bulks. Then you don't have to worry as much about zero counts and you can use something like edgeR's TMM normalization.

All of this is a shame, because the deconvolution method is pretty clever if I do say so myself. I'm quite fond of it, but it is what it is.

Anyway, the block= in center_size_factors() does something different; you can see the detailed documentation for the corresponding C++ function. And of course, you are free to run analyze() on each cluster, but this should be motivated by a desire to look at further structure within each cluster; it does not need to be justified by any improvement in normalization.