Deconvolution for the simulated mouse embryo data

Data preprocessing

We generate simulated data using the get_simulated_data function. For data simulation, users must set the distance parameter to specify the radius for each simulated spot. By setting the label_tag parameter (adata.obs[label_tag]), our simulator can produce both the ground truth deconvolution (cell type proportion) matrix and the simulated spot labels, which represent the cell type of the highest proportion.

import stSCI
import scanpy as sc
from stSCI.utils.simulate import get_simulated_data

adata = sc.read('data/mouse_embryo/embryo_e9.5_slideseq.h5ad')
sim_adata = get_simulated_data(adata, distance=50, label_tag='cell_type')
sc.pl.spatial(sim_adata, color='cluster', spot_size=50)

>>> INFO: Generate simulated ST data, simulated data shape (1368, 27554), average 10.788 cells in each spot. (1.19s).

sc_adata = adata.copy()
st_adata = sim_adata.copy()

sc.pp.highly_variable_genes(sc_adata, flavor="seurat_v3", n_top_genes=6000)
sc.pp.highly_variable_genes(st_adata, flavor="seurat_v3", n_top_genes=6000)
sc.pp.normalize_total(sc_adata, target_sum=1e4)
sc.pp.normalize_total(st_adata, target_sum=1e4)
sc.pp.log1p(sc_adata)
sc.pp.log1p(st_adata)

Model training

sc_adata, st_adata = stSCI.train(sc_adata, st_adata, hvg_count=6000, cluster_key='cell_type', deconvolution=True)

>>> INFO: Filtered 2235 genes.
>>> INFO: Find 3765 same HVGs, result data shapes: [(14758, 3765), (1368, 3765)].
>>> INFO: Generate 9576 edges, 6.000 edges per spot (0.01s).
>>> INFO: Finish PCA (0.49s).
>>> INFO: Finish centroid generation (0.50s).
>>> INFO: Finish centroid generation (0.13s).

>>> Train stSCI: 100%|██████████| 500/500 [00:42<00:00, 11.68it/s]

>>> INFO: Generate 58840 MNN pairs, 3.987 pairs per SC cell; 43.012 pairs per ST spot (0.18s).
>>> INFO: Finish generate deconvolution result (0.07s).
>>> INFO: Train stSCI model (46.53s).

Evaluation

We provide the function plot_deconvolution to visualize the proportion for each cell.

from stSCI.utils.plot import plot_deconvolution
plot_deconvolution(st_adata, spot_size=50)

>>> Predicted deconvolution result

>>> Ground truth deconvolution result

We then further quantify the deconvolution result by the overall PCC, SSIM, RMSE, and JSD.

import st_datasets as stds
pcc_score = stds.metrics.cal_PCC(st_adata.obsm['deconvolution_result'].to_numpy(),
    st_adata.obsm['decon_result'][st_adata.obsm['deconvolution_result'].columns].to_numpy())
ssim_score = stds.metrics.cal_SSIM(st_adata.obsm['deconvolution_result'].to_numpy(),
    st_adata.obsm['decon_result'][st_adata.obsm['deconvolution_result'].columns].to_numpy())
rmse_score = stds.metrics.cal_RMSE(st_adata.obsm['deconvolution_result'].to_numpy(),
    st_adata.obsm['decon_result'][st_adata.obsm['deconvolution_result'].columns].to_numpy())
jsd_score = stds.metrics.cal_JSD(st_adata.obsm['deconvolution_result'].to_numpy(),
    st_adata.obsm['decon_result'][st_adata.obsm['deconvolution_result'].columns].to_numpy())
print(f"PCC: {pcc_score:.3f},", f"SSIM: {ssim_score:.3f},", f"RMSE: {rmse_score:.3f},", f"JSD: {jsd_score:.3f},")

PCC: 0.788, SSIM: 0.632, RMSE: 0.110, JSD: 0.539,