Model request: scPhere (hyperspherical / hyperbolic latent embedding for scRNA-seq)

[nataliaMalicka]

Is your feature request related to a problem? Please describe.
Many core scvi-tools VAE models, including SCVI, use Euclidean latent variables with Gaussian priors. For datasets with many cell types, multi-level batch effects, or hierarchical/continuous structure, very low-dimensional Euclidean embeddings can suffer from crowding and distortion, making interpretation and inter-cluster relationships less reliable when the goal is a 2D/3D representation. I am not aware of a scvi-tools model whose primary latent space is non-Euclidean (hyperspherical or hyperbolic).

Describe the solution you'd like
I would like to add scPhere, a deep generative VAE that embeds scRNA-seq profiles on hyperspheres or in hyperbolic space rather than a Euclidean Gaussian latent. It was designed to reduce crowding/distortion, jointly model multiple batch/confounding vectors, and give interpretable low-dimensional visualizations.

• Paper: Ding & Regev, Nature Communications (2021). https://www.nature.com/articles/s41467-021-22851-4
• Reference implementation: https://github.com/Ding-Group/scPhere

Additional context
I'm part of Jiarui Ding's research group (the lab behind scPhere), and will keep the implementation faithful to the original method. I'm interested in implementing this and opening a PR. Would this be of interest to add to scvi-tools?

Thanks!
Natalie Malicka

[ori-kron-wis]

Hey @nataliaMalicka

That sounds great!
Several tips:

• Please put it under external and use the model-module design principle (as in the reference implementation, it is compact right now). Start with the inheritance of the base model, which brings core scvi-tools functionality.
• Try to reuse as much code as possible (e.g : scvi decoder likely can be used, Anndata manager, trainer- UnsupervisedTrainingMixin and trainingPlan), but for mixins, be aware that they might assume Euclidean-like posterior distributions (mostly Normal, e.g, get_latent() from VAEmixin), so a local implementation of some of the functions is needed, while others might work. Add proper unit testing for all of that (inc core functionality).
• Do not change scvi-tools core functions unless it's a very minor change and code reuse can be done w/o affect other things. Same thing for other models' code - no reason to change them.
• Add the new distributions (hyperspherical/hyperbolic etc ) to scvi/distributions/
• keep other utility functions (data generator, plot, and so on) under a subfolder of util: scvi/external/scphere/utils/
• Try to extend scphere functionality by adding other mixins that we have in scvi-tools but not in the current implementation (DE, DA, and so on) and test them (even better to show on tutorial)
• Add a changelog and list additional required packages - are they really mandatory, or can they be replaced/ manage without them?
• Add docs: model description, end-to-end real-life data tutorial, and manuscript reference (most important as otherwise users will not know about it and how to use it)

@canergen Any other comments?