STING - Spatial Transcriptomics cluster Inference with Nested GNNs

STING incorporates both graphs generated from the spatial proximity of tissue locations (or spots) and spot-specific graphs for related genes. This feature allows STING to better distinguish between clusters and identify meaningful gene-gene relations for knowledge discovery. It is a nested GNN framework that simultaneously models gene-gene and spatial relations. Using the gene expression, we generate a spot-specific gene-gene co-expression graph. We implement an inner GNN for these graphs to generate embeddings for each location. Next, we utilize these embeddings as features in a sample-wide graph generated using spatial information. We implement an outer GNN for this graph to reconstruct the original gene expression data. Finally, STING is trained end-to-end to generate embeddings that capture gene-gene and spatial information, which we input to a clustering algorithm to produce the spatial clusters. Experiments for 45 samples across 9 datasets and 5 spatial sequencing technologies show that STING outperforms the existing state-of-the-art techniques.

Requirements

The model requires the following modules and versions -

PyTorch >= 2.0.1

PyTorch Geometric >= 2.3.1

scanpy >= 1.9.1

Numpy < 2.0.0

Python Optimal Transport >= 0.9.1

Scikit-learn >= 1.2.2

We suggest generating an environment (such as conda) to run the code. You can use the given requirements.txt file to create a new environment. If it does not work, you can create the required conda environment directly by running the following lines sequentially in the shell.

conda create --name <env_name> python==3.11
conda activate <env_name>
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install torch_geometric
pip install scanpy
pip install POT
pip install "numpy<2"

Downloading and Using Datasets

We use .h5ad files for our model. Annotated datasets used in the paper are available for download here (MERMB), here (10xHK), and here (all others). The datasets, their ID and other details as displayed in the manuscript are tabulated below.

Dataset ID	Technology	Sample Count	Cluster Count	Spot Count	Gene Count	Species	Tissue
10xHK	10x Visium	4	4, 5	1,949 - 4,860	9,943	Human	Kidney
10xHB	10x Visium	1	20	3,798	36,601	Human	Breast
10xHPC	10x Visium	12	5, 7	3,460 - 4,789	33,538	Human	Prefrontal Cortex
STAR1	STARmap	3	4	1,049 - 1,088	166	Mouse	Prefrontal Cortex
STAR2	STARmap	1	7	1,207	1,020	Mouse	Prefrontal Cortex
BAR	BaristaSeq	3	6	1,525 - 2,042	79	Mouse	Primary Cortex
OSM	osmFISH	1	11	4,839	33	Mouse	Somatosensory Cortex
MERMPH	MERFISH	5	8	5,488 - 5,926	155	Mouse	Preoptic Hypothalamus
MERMB	MERFISH	15 (chosen from a total 245 samples)	5 - 12	6,000	1,122	Mouse	Brain

To use your own spatial transcriptomics file, use anndata to read the file and store it in an anndata object. The anndata object (named `adata` in the tutorial) should be formatted with `adata.X` containing the raw counts and `adata.obsm['spatial']` containing the spatial locations of the spots.

Tutorial

To see how to use STING to generate embeddings and attention scores, please refer to tutorial.ipynb.

Bugs and Suggestions

Please report any bugs, problems, suggestions, or requests as a Github issue.