Create an atlas landmark

In this example, we will create an atlas landmark, i.e. a compressed approximation of a cell atlas. northstar can either average all cells within each cell type or take a subsample that is by default balanced across cell types, independently on their original abundance:

import anndata
import northstar

# Read in the atlas with annotations
atlas_full = anndata.read_loom('...')

# Make sure the 'CellType' column is set
# if it has another name, rename it
atlas_full.obs['CellType'] = atlas_full.obs['cluster'].astype(str)

# Create a landmark of averages
atlas_ave = northstar.average_atlas(
    atlas_full,
)

# Create a landmark of subsamples
atlas_sub = northstar.subsample_atlas(
    atlas_full,
    n_cells=20,
)

Note

Creating average landmarks does not normalize the cells to one another before averaging. This choice reproduces the null expectation of taking all sequencing reads from those cells and lumping them together. This is not wrong but might be a little naive for some situations. You are free to normalize your atlas before using average_atlas to weight each cell to your liking (e.g. counts per millions). subsample_atlas does not suffer from this ambiguity.