d-scIGM: deep Interpretable Generative Model called for single-cell data analysis

The full description of d-scIGM and its application on published single cell RNA-seq datasets are available.

Download archive with preprocessed data at: https://drive.google.com/drive/folders/1XC4Rhm-c6tMgjPJMCj1FpzW-NqTnM1Fz.

The repository includes detailed installation instructions and requirements, scripts and demos.

1 Schematic overview of d-scIGM and applications.

(a) The overview of d-scIGM, which consists of a PAS encoder and a SC decoder. Given scRNA-seq data as input, the PAS encoder maps it to a low-dimensional latent space. The SC decoder learns topic embeddings and gene embeddings during reconstruction, which ensures the interpretability of the model. (b) The SC decoder’s network architecture. $\boldsymbol{\alpha}^{(t)}$ is the embedding of the $t$-th layer, and the embedding of different layers is mapped to a shared embedding layer through the SC technology, thereby coupling the relationship between different hidden layers. (c) Schematic representation of different applications of d-scIGM.

2 Requirements

• Linux/UNIX/Windows system
• Python >= 3.8
• torch == 1.8.1
• scanpy == 1.9.1

3 Usage

Data format

d-scIGM requires cell-by-cell-gene matrix and cell type information to be entered in .csv object format.

Training

cd model
python main.py

We provide default data for users to understand and debug d-scIGM code.

evaluation

We provide tutorial as shown in directory tutorial/{Cell representation,Pathway enrichment,Time-trajectory inference, and Survival analysis} for introducing the usage of d-scIGM and reproducing the main result of our paper.

Reference

If you use d-scIGM in your work, please cite