romain-lopez
commented
1 year ago Yes! I have used the energy distance function from this package
https://dcor.readthedocs.io/en/latest/functions/dcor.energy_distance.html
This must be calculated in some low-dimensional space, so for example scVI or PCA, between the control cells, and the cells from each guide. The energy distance can also be substituted by the MMD with a linear kernel. Both of these metrics quantify changes between distribution, a useful quantity for assessing impact of those perturbations! Thanks!
In section 5.1 of your paper you explain how to preprocess and split the Norman dataset. My main question regarding this is: Can you further elaborate which interventions (top 30) you selcected with the most significant effect on gene expression for the held-out set? For reproducibility and comparing with your results it would be very helpful to either provide these as a fixed list (of names) or a code snippet to reproduce the same split.
Any help on this topic would be very much appreciated and let me just say I enjoyed reading your paper very much as well!