Index

04-In-situ-spatial-niches_and_domain

Statistical analysis

Based on cell neighborhood and transcriptome proximity, several packages help to identifies niches and spatial domains to highlight structural organization of cells within the tissue. Niches are extremely interesting to define for instance heterogeneous immune cells organisation in cancer sample while spatial domain notion rely more on anatomical region. We focus here on 2 different packages (ie. cellcharter and novae) allowing the definition of both aspect adding metadata to AnnData .obs table. We also explain how to import anatomical regions coming from xenium explorer lasso shape .csv export.

Module specifications

Input data	Computational environnement	Output data
AnnData	spatial in-situ imaging-based Python	AnnData

Code

This module rely on adding niches and spatial domains to AnnData metadata. It relies on two python packages:

Cell Charter

Varonne M. et al. CellCharter reveals spatial cell niches associated with tissue remodeling and cell plasticity. Nature Genetics volume 56, pages 74–84 (2024)
* Article
* Documentation

Novae

Blampey Q. et al. Novae: A Graph-Based Foundation Model for Spatial Transcriptomics Data. Mol Syst Biol. 2019;15(6):e8746. Published 2019 Jun 19. doi:10.15252/msb.20188746
* Article
* Documentation

01. Cell Charter niches analysis

The code below is based on the following CellCharter tutorial notebook.

import cellcharter as cc
import scispy as scis
import torch
import scvi

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print('torch version is {}, device is {}'.format(torch.__version__, device))
print('scanpy version is {}, anndata version is {}'.format(sc.__version__, ad.__version__))
print('squidpy version is {}'.format(sq.__version__))
print('scvi version is {}'.format(scvi.__version__))
print('cellcharter version is {}'.format(cc.__version__))

adata = ad.read_h5ad('000-outs/ad6_scanvi_scmusk.h5ad')
adata.X = adata.layers["counts"].copy()
sc.pp.normalize_total(adata)
sc.pp.log1p(adata)

scvi.settings.seed = 12345
scvi.model.SCVI.setup_anndata(
    adata, 
    layer="counts", 
    batch_key='sample'
)

model = scvi.model.SCVI(adata)
model.train(early_stopping=True, enable_progress_bar=True) 
adata.obsm['X_scVI'] = model.get_latent_representation(adata).astype(np.float32)

sq.gr.spatial_neighbors(adata, library_key='sample', coord_type='generic', delaunay=True, spatial_key='spatial')
cc.gr.remove_long_links(adata)
cc.gr.aggregate_neighbors(adata, n_layers=3, use_rep='X_scVI', out_key='X_cellcharter', sample_key='sample')

gmm = cc.tl.ClusterAutoK(
    n_clusters=(10,30), 
    max_runs=10, 
    model_params=dict(random_state=12345, trainer_params=dict(accelerator='gpu', devices=1))
)
gmm.fit(adata, use_rep='X_cellcharter')

adata.obs['cc_niches'] = gmm.predict(adata, use_rep='X_cellcharter')

adata.write('000-outs/ad6_scanvi_scmusk_cc.h5ad')

sub = adata[adata.obs['sample'] == '2658']
sq.pl.spatial_scatter(sub, color="cc_niches", shape=None, size=0.1)

02. Novae spatial domains analysis

import anndata as ad
import scanpy as sc
import novae

adata = ad.read_h5ad('000-outs/ad6_scanvi_scmusk_cc.h5ad')
novae.utils.spatial_neighbors(adata, slide_key='sample', radius=80)
novae.plot.connectivities(adata)

# select foundation model to use, here mouse brain
model = novae.Novae.from_pretrained("MICS-Lab/novae-brain-0")

# Option 1: zero-shot
model.compute_representations(adata, zero_shot=True)

# Option 2: fine-tuning
#model.fine_tune(adata, accelerator='cuda')
#model.save_pretrained(save_directory="./my-model-directory")

model.compute_representations(adata, accelerator='cuda')

# 7 spatial domains by default, here asking for 15
model.assign_domains(adata, level=15)

novae.plot.domains(adata)
adata.write('000-outs/ad6_scanvi_scmusk_cc_novae.h5ad')

# novae.plot.pathway_scores(adata[adata.obs['sample'] == 'C31'], pathways="h.all.v2024.1.Hs.json", figsize=(10, 7))
# novae.plot.pathway_scores(
#    adata, pathways="h.all.v2024.1.Hs.json", figsize=(4, 4), slide_name_key="sample", pathway_name='HALLMARK_COMPLEMENT',
# )