Assume I have a dataset with cells from many patients. I want to create a space where each observation is the average of all the cells of a certain type in one patient (i.e., the number of observations will be the number of cell types X the number of patients).

What will be the most intuitive way to summarize the latent space for the new observations?

- Simply average the scVI latent space coordinates of all the cells in each observation to generate a latent coordinate for the observation?
- Same, but weight average by posterior probabilities of cells in the observation?
- Mode of posterior distribution over the cells in the observation?
- (other?)