ST data presents challenges such as uneven cell density distribution, low sampling rates, and complex spatial structures. Traditional spot-based analysis strategies struggle to effectively address these issues. STMiner explores ST data by leveraging the spatial distribution of genes, thus avoiding the biases that these conditions can introduce into the results.
Here we propose “STMiner”. The three key steps of analyzing ST data in STMiner are depicted.
(Left top) STMiner first utilizes Gaussian Mixture Models (GMMs) to represent the spatial distribution of each gene and the overall spatial distribution. (Left bottom) STMiner then identifies spatially variable genes by calculating the cost that transfers the overall spatial distribution to gene spatial distribution. Genes with high costs exhibit significant spatial variation, meaning their expression patterns differ considerably across different regions of the tissue. The distance array is built between SVGs in the same way, genes with similar spatial structures have a low cost to transport to each other, and vice versa. (Right) The distance array is embedded into a low-dimensional space by Multidimensional Scaling, allowing for clustering genes with similar spatial expression patterns into distinct functional gene sets and getting their spatial structure.
Please visit STMiner Documents for installation and detail usage.
from STMiner import SPFinder
You can download the demo dataset from GEO, or you can also download them from STMOMICS. STMiner can read spatial transcriptome data in various formats, such as gem, bmk, and h5ad (see STMiner Documents). We recommend using the h5ad format, as it is currently the most widely used and supported by most algorithms and software in the spatial transcriptomics field.
sp = SPFinder()
file_path = 'Path/to/your/h5ad/file'
sp.read_h5ad(file=file_path, bin_size=1)
The parameter bin_size specifies the size of merged cells (spots). If not specified, no merging is performed. If set to 50, 50x50 cells/spots will be merged into a single cell/spot. Due to low sequencing depth in some datasets, cells/spots are often merged during analysis (e.g., stereo-seq). However, 10x data typically does not require merging.
sp.get_genes_csr_array(min_cells=500, log1p=False)
sp.spatial_high_variable_genes()
You can check the distance of each gene by:
sp.global_distance
Gene | Distance |
---|---|
geneA | 9998 |
geneB | 9994 |
... | ... |
geneC | 8724 |
The first column is the gene name, and the second column is the difference between the spatial distribution of the gene and the background.
A larger difference indicates a more pronounced spatial pattern of the gene.
sp.fit_pattern(n_comp=20, gene_list=list(sp.global_distance[:1000]['Gene']))
n_comp=20 means each GMM model has 20 components.
sp.build_distance_array()
sp.cluster_gene(n_clusters=6, mds_components=20)
The result is stored in genes_labels:
sp.genes_labels
The output looks like the following:
gene_id | labels | |
---|---|---|
0 | Cldn5 | 2 |
1 | Fyco1 | 2 |
2 | Pmepa1 | 2 |
3 | Arhgap5 | 0 |
4 | Apc | 5 |
.. | ... | ... |
95 | Cyp2a5 | 0 |
96 | X5730403I07Rik | 0 |
97 | Ltbp2 | 2 |
98 | Rbp4 | 4 |
99 | Hist1h1e | 4 |
import seaborn as sns
sns.clustermap(sp.genes_distance_array)
Note: To better visualization, images need cutting border of the original dataset. Anyhow, you can download the processed image here.
sp.get_pattern_array(vote_rate=0.3)
img_path = 'path/to/downloaded/image'
sp.plot.plot_pattern(vmax=99,
heatmap=False,
s=5,
reverse_y=True,
reverse_x=True,
image_path=img_path,
rotate_img=True,
k=4,
aspect=0.55)
sp.plot.plot_intersection(pattern_list=[0, 1],
image_path=img_path,
reverse_y=True,
reverse_x=True,
aspect=0.55,
s=20)
sp.plot.plot_genes(label=0, vmax=99)
Attribute | Type | Description |
---|---|---|
adata | Anndata | Anndata for loaded spatial data |
global_distance | pd.DataFrame | OT distance between gene and background |
genes_labels | pd.DataFrame | Gene name and their pattern labels |
genes_patterns | dict | GMM model for each gene |
genes_distance_array | pd.DataFrame | Distance between each GMM |
kmeans_fit_result | obj | Result of k-means |
mds_features | pd.DataFrame | embedding features after MDS |
- Peisen Sun ([email protected])
- Kai Ye ([email protected])