Scalable optimal transport-based method for multimodal integration of RNA-seq and ATAC-seq data.

pip install scSAGA

We recommend creating a conda environment first:

conda env create -n scmint -f environment.yml
conda activate scmint
pip install -e .

Note on PyTorch: pip install scSAGA will install a CPU-only version of PyTorch by default. If you need GPU support, install PyTorch manually with the appropriate CUDA version from pytorch.org before installing scSAGA.

Once installed, run the analysis from the command line:

scsaga config/input.yml

Create a YAML config file specifying your datasets and parameters. A template is provided in config/input.yml. More datasets can be added in the same format as needed.

anchor: "rna1"

datasets:
  - name: "rna1"
    modality: "rna"
    counts: "/path/to/rna_normalized_counts.mtx"
    barcodes: "/path/to/rna_barcodes.txt"
    features: "/path/to/rna_features.txt"
    pca: "/path/to/rna_pca_50.txt"

  - name: "atac"
    modality: "atac"
    counts: "/path/to/atac_normalized_counts.mtx"
    barcodes: "/path/to/atac_barcodes.txt"
    features: "/path/to/atac_features.txt"
    pca: "/path/to/atac_pca_50.txt"

  # Add more datasets as needed:
  # - name: "rna2"
  #   modality: "rna"
  #   ...

output_dir: "/path/to/output_directory"

# sketch_size:   # Optional: downsample cells via geometric sketching

# --- Hyperparameters ---
s_shared_cells:   # Estimated number of shared cells across modalities
M_samples:        # Anchor pairs sampled per OT iteration
alpha:             # Update step size (0 to 1)
S_iterations:        # Number of SAGA iterations
gw_epsilon:      # Convergence threshold
gw_reg:            # Sinkhorn regularization strength

Results are saved to the directory specified by output_dir:

• T_<dataset>_to_<anchor>.npy — transport plan for each dataset pair
• joint_embedding_2d.png — PCA plot of the joint embedding
• joint_embedding_2d.csv — 2D coordinates for the joint embedding
• saga_runtimes.txt — timing breakdown and alignment scores

After cloning the repo:

git clone https://github.com/Swethasree-Bhattaram/scSAGA.git
cd scSAGA
conda env create -n scmint -f environment.yml
conda activate scmint
pip install -e .

The -e flag installs in editable mode, so changes to the source code take effect immediately without reinstalling.