Analysis on number of control samples required

[emdann]

Does the PAC design allow to use less control samples compared to PC design?

Plot AUPRC (or other metric) as a function of number of control donors for PC and PAC design (from 3 to 16 control donors). Take 3 random initialization for each number of samples, esp the lower numbers, to assess variance.

[emdann]

Optimizing

• only 5 query samples (to always have more ctrl than query samples)
• Select query samples that have most of the perturb_pop cells (otherwise high variability in cell numbers leads to high variance between different initializations)