Inference of spatial chromatin accessibility via integration of spatial transcriptomics and single-cell multiomics data.

ISON (Integrated Spatial Omics Network) is a computational tool developed to infer spatial chromatin accessibility by integrating spatial transcriptomics with single-cell multiomics data. It performs dimension reduction and learns the embeddings in a joint latent space and uses these embeddings to predict chromatin accessibility data in a spatial context.

Input

• Single-cell multiome
  • scRNA (cells x genes)
  • scATAC (cells x peaks)
• Spatial transcriptomics (spots x genes)
• Coordinates file with only x and y coordinates (spots x 2)

Installation

• Python >= 3.8 The environment.yml file lists all the libraries required to run ISON, with PyTorch configured for CPU-only usage. If you have a GPU and wish to enable CUDA support, please install the appropriate version of PyTorch by following the instructions at https://pytorch.org/get-started/locally.

conda env create -f environment.yml
conda activate ison_env

Running the model

To run ISON, run the run-ison.py .

python run-ison.py --scRNA "$SCRNA" \
                        --scATAC "$SCATAC" \
                        --ST "$ST" \
                        --coords "$COORDS" \
                        --output-dir "$OUTPUT_DIR" \
                        --lambda1 "$LAMBDA1" \
                        --lambda2 "$LAMBDA2" \
                        --K "$K" \
                        --batch_size "$BATCH_SIZE"\

The default parameters are set as: $\lambda_1$ = 15, $\lambda_2$ = 0.001, $K$ = 8 and batch_size = 512. To evaluate ISON, there are toy datasets provided in the data folder in .h5ad format. evaluation.py in project folder is provided to compute the Pearson correlation coefficient (PCC) between the true and predicted ATAC values.

The output of ISON is spatial chromatin accessibility, denoised spatial gene expression and the non-negative weights.

Tutorial

To run ISON using processed P21 data, run the tutorial.sh script. Processed P21 and P22 data is available at https://drive.google.com/drive/folders/15LMxbizMdrBuPALeMwLvaeg5yc63zljh?usp=share_link. The evaluation.py script computes the PCC between ground truth and predicted chromatin accessibility values for ISON's performance evaluation.

Make sure the .sh file is executable, then run it from your terminal:

chmod +x tutorial.sh
./tutorial.sh

Choosing Optimal Hyperparameters

The choice of hyperparameters depends on both prior biological knowledge and the desired level of resolution in your analysis.

• $K$: If the expected number of cell types is known, $K$ can be set accordingly. Otherwise, treat $K$ as a resolution parameter: larger values enable the model to distinguish finer-grained cell populations, while smaller values are better suited for capturing broader cell type groupings.
• $\lambda_1$: To select an appropriate value for $\lambda_1$, refer to the lambda1_tune Jupyter notebook. This notebook implements a biologically informed tuning approach that uses peak–gene distance information to guide parameter selection.