scANVI

[sjfleming]

Implement the scANVI model
https://docs.scvi-tools.org/en/stable/user_guide/models/scanvi.html

scANVI is a super cool model for cell type annotation that was published by Xu et al in 2021, and is based on the M1+M2 model from Kingma's 2014 "semi-supervised deep learning with deep generative models".

This implementation re-uses as much of the scVI implementation as possible.

In the future it may be possible to extend this in two very interesting ways for CAS:

1. calibrate the uncertainties in cell type label prediction to be accurate
2. compute an estimate for out-of-distribution probability, i.e. the probability that an observation is "novel"