Reconstructing spatial coordination for mouse brain cortex

Data preprocessing

We load the SC and ST data from the squidpy package.

import stSCI
import scanpy as sc
import squidpy as sq

sc_adata = sq.datasets.sc_mouse_cortex()
sc_adata.var_names_make_unique()
sc_adata.obs_names_make_unique()
sc.pp.normalize_total(sc_adata, target_sum=1e4)
sc.pp.log1p(sc_adata)

For ST data, we selected the cortex part from the mouse brain for integrating with the SC data.

st_adata = sq.datasets.visium_fluo_adata()
st_adata = st_adata[[
    'Cortex' == label.split('_')[0]
    for label in st_adata.obs['cluster']
]]
st_adata.var_names_make_unique()
st_adata.obs_names_make_unique()
sc.pp.normalize_total(st_adata, target_sum=1e4)
sc.pp.log1p(st_adata)

Model training

For reconstruct spatial coordination, set the coor_reconstruction parameter to True to generate the predicted coordination for SC data. The predicted coordination will save at sc_adata.obsm['spatial'].

sc_adata, st_adata = stSCI.train(sc_adata, st_adata, cluster_key='cell_subclass', coor_reconstruction=True)

>>> INFO: Filtered 1890 genes.
>>> INFO: Find 1110 same HVGs, result data shapes: [(21697, 1110), (1259, 1110)].
>>> INFO: Generate 8813 edges, 6.000 edges per spot (0.01s).
>>> INFO: Finish PCA (0.43s).
>>> INFO: Finish centroid generation (1.20s).
>>> INFO: Finish centroid generation (0.02s).

>>> Train stSCI: 100%|██████████| 500/500 [00:50<00:00,  9.99it/s]

>>> INFO: Generate 55344 MNN pairs, 2.551 pairs per SC cell; 43.959 pairs per ST spot (0.26s).
>>> Set distance threshold to 278.820.

>>> INFO: Train SC coordination: 100%|██████████| 21697/21697 [00:07<00:00, 2942.93it/s]

>>> INFO: Train SC coordination (8.09s).
>>> INFO: Train stSCI model (67.26s).

Visualization

sc.pl.spatial(sc_adata, color='cell_subclass', spot_size=120)