Vamp prior testing

[Hrovatin]

Things to check:
How pseudoinputs move accross epochs:

cross_system_integration/notebooks/tryout/xxjoint_vampprior.py

Line 765 in 58a3642

# ### VAMP - change of pseudoinputs over training

Things that may help integration:

• Different proportions of 2 systems (e.g. if we have prior components only from one system this may force other to be closer)
• Do prior components of the two systems migrate closer to each other during training?

Notes on analysis:

• Would use only cell types that are well represented in both species
• Balancing of N components per ct may be useful so that in all 3 settings (priors from both systems or one of the two systems) have same number of components per ct.

Notes on VAMP:

• Takes as pseudoinputs real data (expression, covar) and embeds them to get prior components.
• The pseudoinputs change during training, leading to their embed (prior components) also changing,