sceptre default covariate model

[redbybeing]

Hi,
I have a naive question for the default model formula (for high moi) used in sceptre.
formula(~ log(response_n_nonzero) + log(response_n_umis) + log(grna_n_nonzero) + log(grna_n_umis) + response_p_mito + batch)
I am curious why log is applied for the count values? I happened to try removing log from the formula and run and it did seem to increase the number of false positives in the negative control pairs. So I guess somehow log is a must for better calibration?

And, is the exact same formula used for all pairs tested? Would there be a case where some pairs are better fitted with another model?

Thank you!

Best,
Jiseok

[ekatsevi]

I am curious why log is applied for the count values?

Our rationale is primarily for the sequencing depth variables like response_n_umis. To see why the logarithmic transformation makes sense at least for this variable, see Section 5.2.2 of my lecture notes. This section is in the context of Poisson regression, but it's similar for NB regression. Furthermore, the notes discuss adding a term like log(response_n_umis) as an offset (i.e. with coefficient 1), but in sceptre we add it as a covariate (i.e. with a coefficient we fit based on the data) for added robustness.

And, is the exact same formula used for all pairs tested? Would there be a case where some pairs are better fitted with another model?

Yes, the same formula is used for all pairs tested. We have not come across a case yet where we need to use different formulas for different pairs.

[redbybeing]

Thank you so much for the kind explanations Eugene!