I have two questions related to scanvi and scvi model options:
Using scanvi for Predicting Unlabeled Cells:
I’ve trained a scvi model on 1 million cells with cell type annotations. Now, I have a proportion of unlabeled cells and would like to predict their cell types using scanvi. After initializing the scanvi model, should I use scanvi_model.predict() to achieve this?
Understanding Model Parameters:
In the scvi model initialization, I used the following parameters:
python
model = scvi.model.SCVI(adata, n_layers=2, n_latent=30, gene_likelihood=“nb”)
I’m curious about the impact of adjusting n_latent and gene_likelihood. Could you explain how varying these parameters affects the output of the learning model? For example, using different values for n_latent and choosing a different gene_likelihood.
Here’s the code snippet I used for setting up my data:
thanks alot for your time and help
adata=sc.read_h5ad(“/pasteur/zeus/projets/p02/LabExMI/singleCell/V3/scRNA_NS_IAV_COV/results/merged_object/adata.h5ad”)
adata=adata[:,highly_variable]
scvi.model.SCVI.setup_anndata(adata, layer = “counts”,
categorical_covariate_keys=[“library”],
continuous_covariate_keys=[‘pct_counts_mt’, ‘total_counts’, “pct_counts_RP”])
model.train()
adata.obsm[‘X_scVI’] = model.get_latent_representation()
adata.layers[‘scvi_normalized’] = model.get_normalized_expression()
scanvi_model = scvi.model.SCANVI.from_scvi_model(
model,
adata=adata,
labels_key=“cell_type”,
unlabeled_category=“unassigned”,
)
adata.obsm[SCANVI_LATENT_KEY] = scanvi_model.get_latent_representation(adata)