Gaussian Smoothing

This simple case shows how to run the function of Gaussian smoothing in Stereopy. Gaussian smoothing can make expression matrix closer to reality [Shen22].

Generally, preprocessing, including quality control, filtering and normalization, and PCA should be done before running Gaussian smoothing. Meanwhile, you also have to save raw expression matrix by data.tl.raw_checkpoint before the operations which will change the expression matrix.

Observation on genes

import stereo as st
import warnings
warnings.filterwarnings('ignore')

# read data
input_file = './SS200000141TL_B5_raw.h5ad'
data = st.io.read_ann_h5ad(input_file, spatial_key='spatial')

# preprocessing
data.tl.cal_qc()
data.tl.filter_cells(min_gene=300, pct_counts_mt=10)
data.tl.filter_genes(min_cell=10)

# save raw data
data.tl.raw_checkpoint()

data.tl.normalize_total(target_sum=10000)
data.tl.log1p()

# PCA
data.tl.pca(use_highly_genes=False, n_pcs=50, svd_solver='arpack')

# Gaussian smoothing
data.tl.gaussian_smooth(n_neighbors=10, smooth_threshold=90)
data.tl.scale(max_value=10) # only for gaussian_smooth_scatter_by_gene

# plot
data.plt.gaussian_smooth_scatter_by_gene(gene_name='C1ql2')
data.plt.gaussian_smooth_scatter_by_gene(gene_name='Irx2')
data.plt.gaussian_smooth_scatter_by_gene(gene_name='Calb1')

Obervation on clusters

Whereafter, if you want to do other operations such as clustering, the same processing steps should have been done.

import stereo as st
import os
import warnings
warnings.filterwarnings('ignore')

# read data
input_file = './SS200000141TL_B5_raw.h5ad'
data = st.io.read_ann_h5ad(input_file, spatial_key='spatial')

# preprocessing
data.tl.cal_qc()
data.tl.filter_cells(min_gene=300, pct_counts_mt=10)
data.tl.filter_genes(min_cell=10)

data.tl.raw_checkpoint() # save raw data

data.tl.normalize_total(target_sum=10000)
data.tl.log1p()

# PCA
data.tl.pca(use_highly_genes=False, n_pcs=50, svd_solver='arpack')

# Gaussian smoothing
data.tl.gaussian_smooth(n_neighbors=10, smooth_threshold=90)
data.tl.normalize_total(target_sum=10000)
data.tl.log1p()

# clustering
data.tl.pca(use_highly_genes=False, n_pcs=50, svd_solver='arpack')
data.tl.neighbors(pca_res_key='pca', n_pcs=30, res_key='neighbors')
data.tl.leiden(neighbors_res_key='neighbors', res_key='leiden')
data.plt.cluster_scatter(res_key='leiden')

<AxesSubplot:>

Gaussian smoothing is able to make the clustering result to more subtypes as below:

Performance

The example data contains 61857 cells (bins) and 24562 genes.

Machine configuration as below:

physical cores	logic cores	memory
12	48	250G

This module only supports single process.

process	memory(max)	cpu	time
1	8.5G	2736%	1.8m