scVI.jl

A Julia package for fitting VAEs to single-cell data using count distributions. Based on the Python implementation in the scvi-tools package.

Overview

The scVI model was first proposed in Lopez R, Regier J, Cole MB et al. Deep generative modeling for single-cell transcriptomics. Nat Methods 15, 1053-1058 (2018).

More on the much more extensive Python package ecosystem scvi-tools can be found on the website and in the corresponding paper Gayoso A, Lopez R, Xing G. et al. A Python library for probabilistic analysis of single-cell omics data. Nat Biotechnol 40, 163–166 (2022).

This is the documentation for the Julia version implementing basic functionality, including the following (non-exhausive list):

• standard and linearly decoded VAE models
• support for negative binomial generative distribution with and without zero-inflation
• different ways of specifying the dispersion parameter
• store data in a (very basic) Julia version of the Python AnnData objects
• several built-in datasets
• training routines supporting a wide range of customisable hyperparameters

Installation

The package can be downloaded from the Github repo and added with the Julia package manager via

julia> ]
pkg > add "https://github.com/maren-ha/scVI.jl"

or alternatively by

julia> using Pkg; Pkg.add(url="https://github.com/maren-ha/scVI.jl")

Contents

• Data processing
• • AnnData object ind I/O
  • Library size and normalization
  • Filtering
  • Simple transformations
  • Dimension reduction
  • Highly variable genes
  • Loading built-in datasets
• The scVAE model
• • Encoder
  • Decoder
  • VAE model
• The scLDVAE model
• Model training
• Model evaluation
• • Extract latent representations
  • Dimension reduction and plotting
  • Sampling from the trained model