Inferring Causal Trajectories from Spatial Transcriptomics Using CASCAT

CASCAT is a tree-shaped structural causal model with the local Markovian property between clusters and conditional independences to infer a unique cell differentiation trajectory, overcoming Markov equivalence in high-dimensional, non-linear data. CASCAT eliminates redundant links between spatially close but independent cells, creating a causal cell graph that enhances the accuracy of existing spatial clustering algorithms.

Installation & Setup

This step can be finished within a few minutes.

1. Install Miniconda if not already available.
2. Create a new cascat environment, activate it, and install the basic packages.

conda create -n cascat python==3.10 -y 
conda activate cascat

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==2.6.1 pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu118.html
pip install scanpy==1.10.1 matplotlib networkx scikit-misc pydot pot numpy==1.26.4 numba==0.57.1 scikit-learn==1.5.2 
pip install cupy-cuda11x numba==0.60.0 numba-scipy==0.4.0 pandas==2.2.3 scipy==1.11.0

4. (optinal) Install R to generate simulated data.

conda create -n r_env r-essentials r-base -y; 
conda activate r_env
conda install r-mclust
export R_HOME='/home/yourname/miniconda3/envs/r_env/lib/R'
export rScript = '/home/yourname/miniconda3/envs/r_env/bin/Rscript'

Dataset

We provide example dataset tree1 in the ./data/tree1/. Other simulation and real data is hosted on Figshare.

📚 Quick Start

python main.py --YAML ./config/tree1.yml --mode train --verbose True

The output of CASCAT is a new Anndata object data_processed.h5ad under ./result, with the following information stored within it:

• adata.obs['cascat_clusters'] The predicted cluster labels.
• adata.obsm['cascat_embedding'] The generated low-dimensional cell embeddings.
• adata.uns['cascat_connectivities'] The inferred trajecory topology connectivities.
• adata.uns['CMI'] The inferred conditional mutual information matrix for each cluster.

📈 Benchmarking

The YAML files for all datasets are stored on config/yaml/CMI folder, and the comparison method scripts are located in the submodules folder.

🔥 Run CASCAT

To run CASCAT, follow the steps below:

Step1: Data Preparation

CASCAT takes AnnData formatted input, stored in .h5ad files, where obs contains cell/spot information and var holds gene annotations.

To use the data, place it in a folder, then update the adata_file field in the tree1.yml configuration to reflect the relative path to the data.

Step2: Cluster

1. update params in ./config/tree1.yml

   1. CMI_dir as the directory for storing the casual cell graph outputs.
      1. We have accelerated the computation process using GPUs, completing the analysis of 2000 cells within 3 minutes.
      2. We have provided the pre-caculated CMI values between cells in the Google Drive.
   2. percent as the percentage of the causal cell graph to be removed.
      1. default is 0.1 in scRNA-seq dataset and 0.15 in ST dataset.

2. To run CASCAT get cluster result, you can execute following code:

   python main.py --YAML ./config/tree1.yml --mode train --verbose True

   • Note: To access the clustering metrics, set verbose=True and store ground-truth cluster labels in adata.obs['cluster'].

Step3: Trajectory Inference

1. update params in ./config/tree1.yml

   1. emb_path is the path of clustering embedding.
   2. job_dir is the directory of storing the clustering output.
   3. output_dir is the directory of storing the trajectory output.

2. To run CASCAT get only trajectory result, you can execute following code:

   python main.py --YAML ./config/tree1.yml --mode infer

   • Note: To access the TI metrics, store the true pseudo-time labels in adata.uns['timecourse'] and the trajectory topology in adata.uns['milestone_network'].

Step4: Visualization

To visualize the results, refer to the Visualization.ipynb notebook

🎉 InformationMeasures

We've implemented the Python version of InformationMeasures.jl, enhanced with a kernel function.

Consult the InfoMeasure.ipynb for usage details.

In addition, we also provide a GPU version implemented with CuPy, as well as a parallel version implemented with Numba to accelerate the computation of conditional mutual information.