'outer_frac' in liana:::preprocess_scores

[junghpark]

Thanks for the awesome package!

Could you provide further logic on how you usually set the parameter for outer_frac? For instance, if I have 4 samples for 3 conditions (12 samples in total), would 4/12 (=1/3) be a good fraction expecting the samples in the same condition to share a similar set of cells and LR pairs?

Thank you!

[dbdimitrov]

Hi @junghpark,

Indeed, 1/3 should be a good fraction. Another key parameter would be lr_fill, which would essentially dictate how you treat missing values. If left as nan/None, then these values will be imputed during decomposition. If you having zeroes is biologically meaningful (e.g. expressed/not between conditions) then you could also consider filling them as 0s. From my experience, the latter will lead to more pronounced differences between conditions.

Hope this helps!
Daniel

[earmingol]

@junghpark just do complement @dbdimitrov 's answer:

If you are Ok capturing very sample-specific factors, which may be capturing CCC of just a few samples, you can be relaxed and use a low fraction. However, if you want to keep LRs and cell types that may be absent in a few samples but you know that they are present in all the others, then I recommend using higher fractions. Normally I try to use a 0.7-0.8 fraction so I am flexible enough, but knowing that the CCC I capture these are present in most of the samples. Also if you have missing cells/LRs in a few samples, having a fraction like the one I use may also help to predict those missing values as @dbdimitrov indicated (when treated as NaNs).