Downstream Analysis

Latent space

Extract the latent representation and store it in adata.obsm:

adata.obsm["X_spatialvi"] = model.get_latent_representation()

Use RAPIDS for GPU-accelerated downstream analysis:

z = model.get_latent_representation(backend="rapids")  # returns cupy array

Clustering

Apply standard scanpy clustering to the latent space:

import scanpy as sc

sc.pp.neighbors(adata, use_rep="X_spatialvi")
sc.tl.leiden(adata)
sc.pl.embedding(adata, basis="spatial", color="leiden")

Deconvolution (DestVI)

proportions = model.get_proportions()   # DataFrame (n_spots × n_celltypes)
model.plot_cell_type_map(cell_type="Neuron")

Niche prediction (scVIVA)

ct_composition = model.predict_neighborhood()      # (n_cells, n_celltypes)
niche_activation = model.predict_niche_activation()  # (n_cells, n_niches)

Differential expression

de = model.differential_expression(groupby="cell_type", mode="change")
de[de["is_de_fdr_0.05"]].head(20)