The scviR package provides an experimental interface between R and scvi-tools.
Our first release addresses the use of the totalVI model for CITE-seq data.
- The scviR vignette works through a chunk of the colab tutorial for scvi-tools 0.19.0; 0.20.0 employs muon, and this has not been addressed yet.
- scviR defines python infrastructure via the basilisk
discipline; the main python dependencies are declared in
R/basilisk.R. - We have collected a number of intermediate results so that the outputs of totalVI (and other VI procedures) can be explored without taking the time to fit the model. An example in the CITE-seq domain is the anndata instance for the 5k-10k PBMC dataset with representations of the latent space, cluster assignments, and UMAP projection:
> tot = getTotalVI5k10kAdata() # retrieved on first call from Open Storage Network, cached
> tot
AnnData object with n_obs × n_vars = 10849 × 4000
obs: 'n_genes', 'percent_mito', 'n_counts', 'batch', '_scvi_labels', '_scvi_batch', 'leiden_totalVI'
var: 'highly_variable', 'highly_variable_rank', 'means', 'variances', 'variances_norm', 'highly_variable_nbatches'
uns: '_scvi_manager_uuid', '_scvi_uuid', 'hvg', 'leiden', 'log1p', 'neighbors', 'umap'
obsm: 'X_totalVI', 'X_umap', 'denoised_protein', 'protein_expression', 'protein_foreground_prob'
layers: 'counts', 'denoised_rna'
obsp: 'connectivities', 'distances'
> table(tot$obs$batch)
PBMC5k PBMC10k
3994 6855
> dim(xx$obsm$get("X_totalVI")) # cell positions in 20 dimensional latent space
[1] 10849 20
Vignettes in the package show how to populate a Bioconductor SingleCellExperiment with components of this structure to help compare methods employed in the two frameworks.