Interpretable super-resolution dimension reduction of spatial transcriptomics data by DeepFuseNMF

Overview

DeepFuseNMF is based on a multi-modal neural network that takes advantage of the high-dimensionality of transcriptomics data and the super-resolution of image data to achieve interpretable super-resolution dimension reduction. The high-dimensional expression data enable refined functional annotations and the super-resolution image data help to enhance the spatial resolution.

Based on the super-resolution embedding and the reconstruction of gene expressions, DeepFuseNMF can then perform super-resolution downstream analyses, such as spatial domain detection, gene expression recovery, and identification of embedding-associated genes as well as super-resolution cluster-associated genes.

Installation

Please install DeepFuseNMF from pypi with:

pip install DeepFuseNMF

Or clone this repository and use

pip install -e .

in the root of this repository.

Quick start

Prepare your data and run the following command:

python run_DeepFuseNMF.py --config config.json --device 0 --verbose

config.json is a JSON file that contains the paths to the input data and the output directory. The JSON file should look like this:

{
    "settings": {
        "root_path": "your_root_path",
        "project": "your_project_name"
    },
    "sections": [
        {
            "name": "name of section A",
            "image_path": "image path of section A",
            "spot_coord_path": "spot coordinate path of section A",
            "spot_exp_path": "spot expression path of section A"
        },
        {
            "name": "name of section B",
            "image_path": "image path of section B",
            "spot_coord_path": "spot coordinate path of section B",
            "spot_exp_path": "spot expression path of section B"
        }
    ],
    "paras": {
        "scale_factor": [scale_factor_A, scale_factor_B],
        "radius": [radius_A, radius_B],
        "reference": {"name of section B": "name of section A"}
    }
}

All parameters of run_DeepFuseNMF.py are optional. The following are the:

• --config or -c: the path to the configuration file
• --rank or -r: the rank / number of components of the NMF model, default is 20
• --seed or -s: the random seed, default is 123
• --device or -d: the device to run the model, e.g., 0, 1, 2, etc, default is 0
• --visualize: whether to visualize the results, default is True
• --save_score: whether to save the embedding, use --save_score to save all scores
• --save_model: whether to save the model, use --save_model to save the model
• --verbose: whether to print the log information, use --verbose to print the log information

Please refer to the Tutorial/ and configs/README.md for more details.