Spatially Informed Graph Structure Learning Extracts Insights from Spatial Transcriptomics

STAGUE is an unsupervised representation learning model for Spatial Transcriptomics (ST), utilizing spatially informed graph structure learning (GSL) to integrate spatial clustering and cell-cell interaction (CCI) inference tasks into a unified framework. STAGUE can jointly infer biologically meaningful cell representations and cell graph adjacency matrix for ST data. The learned cell representations can be used for spatial clustering, and the adjacency matrix aids in inferring novel CCIs.

Installation & Setup

This step can be finished within a few minutes.

1. Install Miniconda if not already available.

2. Create a new stague environment, activate it, and install the basic packages.

conda create -n stague python==3.10 -y
conda activate stague
pip install scanpy[leiden]==1.9.6 scikit-misc==0.3.0 louvain==0.8.1 networkx==3.2.1 rpy2==3.5.14 matplotlib jupyterlab

3. Install PyTorch and PyG. To select the appropriate versions, you may refer to the official websites of PyTorch and PyG. The following commands are for CUDA 11.8.

pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118
pip install torch_geometric pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu118.html

Datasets

We provide an example dataset of mouse primary visual cortex (V1) under ./data/V1. All the fifteen datasets used in the paper can be accessed from Google drive.

Run STAGUE

STAGUE takes a standard AnnData (adata) object as input. The observations obs are cells/spots and variables var are genes. Some annotations of observations may be stored in adata.obs, such as cell types celltype and true cluster labels cluster. The output of STAGUE is a new adata object, with the following information stored within it:

• adata.obs['cluster_pred'] The predicted cluster labels.
• adata.obsm['embedding'] The inferred low-dimensional cell embeddings.
• adata.obsp['learned_adj_raw'] The learned cell graph adjacency matrix before symmetric normalization.
• adata.obsp['learned_adj_normalized'] The learned cell graph adjacency matrix after symmetric normalization.

To run STAGUE, you can execute the main.py file. The required arguments are:

• --adata_file Path to the input AnnData file, e.g., ./data/V1/raw/adata.h5ad.
• --output_dir Directory path where outputs will be saved, e.g., ./result/V1. A file named adata_processed.h5ad will be generated under this directory.
• --n_clusters Number of clusters to be identified.

Other optional arguments are set with default values. In most situations, running main.py with the default values can yield satisfactory performance. However, if desired, these values can be adjusted to obtain different clustering results. For guidance on hyperparameter adjustment, use the -h option to access an overview and tuning suggestions for key parameters. For example:

• --clu_model Unsupervised clustering algorithms applied to the learned cell embeddings. Choose one from {kmeans, mclust, louvain, leiden}. Before using mclust, ensure that R is installed on your system along with the mclust package. The resolution for louvain and leiden is automatically searched to match n_clusters.
• --adj_weight Weight of the cosine similarity term in the learned cell graph adjacency matrix. Lower values (<= 0.5) are recommended for datasets with low gene coverage.
• --k Number of neighbors for determining the cutoff distance when inferring the learner view's cell graph adjacency matrix. Larger values (>= 15) are preferred.

You can run a demo with the V1 dataset:

python main.py --adata_file ./data/V1/raw/adata.h5ad --output_dir ./result/V1 --n_clusters 7 --adj_weight 0.3 --k 25 --clu_model kmeans

To visualize the detailed inputs and outputs of STAGUE, refer to the Visualization.ipynb notebook.

Spatial Clustering Benchmarking

To reproduce the reported results in the paper, we provide all the related scripts under the ./benchmark_script directory. The benchmark scripts are generally configured with default hyperparameters. Navigate to ./benchmark_script and run the bash script.

Reference

Spatially Informed Graph Structure Learning Extracts Insights from Spatial Transcriptomics. Advanced Science (2024).

Contact

If you have any questions, please feel free to contact us. We will respond as soon as possible.

• Email: [email protected]
• WeChat: akaruin
• WhatsApp: +852-53008094