GemVAE : Graph Enhanced Multi modal Variational Autoencoder

---

To run :

1. Install the package using the command

git clone https://github.com/karanwxliaa/GemVAE.git

2. Unzip all the datasets in the Data dir
3. Install dependancies

pip install requirments.txt

4. Run the .ipynb Tutorials

---

Data

In the /data directory there are several datasets present in the zipped format which need to be extracted before running.

---

Flow Diagram Representation

graph LR
    init[Initialization] --> importTrain[Import train_GEMVAE]
    init --> importUtils[Import Utils Functions]
    init --> importClustering[Import Clustering Functions]

    importUtils --> calSpatialNet[Cal_Spatial_Net]
    importUtils --> statsSpatialNet[Stats_Spatial_Net]
    importUtils --> mclustR[mclust_R]

    importClustering --> clustering[clustering Function]
    clustering --> pca[PCA]
    clustering --> mclust[mclust_R]
    clustering --> leiden[Leiden]
    clustering --> louvain[Louvain]

    importTrain --> trainGEMVAE[train_GEMVAE Function]
    trainGEMVAE --> prepareData[Prepare Data]
    prepareData --> calSpatialNet
    prepareData --> pruneNet[Prune Spatial Net]
    pruneNet --> spatialClustering[Use Clustering]

    trainGEMVAE --> initializeModel[Initialize GEMVAE Model]
    initializeModel --> modelDef[Define GATE Model]
    modelDef --> encoder[Encoder Layers]
    modelDef --> decoder[Decoder Layers]

    initializeModel --> trainingLoop[Training Loop]
    trainingLoop --> epoch[Run Epoch]
    epoch --> updateModel[Update Model Parameters]

    trainingLoop --> inferModel[Infer with Model]
    inferModel --> analysis[Analyze and Visualize Results]

    tutorial[Tutorial Notebook] --> followSteps[Follow Tutorial Steps]
    followSteps --> prepareData
    followSteps --> trainGEMVAE
    followSteps --> inferModel
    followSteps --> analysis

Loading

---

Results

Spatial Multo-Omic clustering of

1. Stereo SiteSeq dataset
2. Landau Spots dataset

---

Tutorials

Tutorial_Landau_BC = Landau Spots Breast Cancer
Tutorial_Landau_SR1 = Landau Spots Spleen Rep 1
Tutorial_Landau_SR2 = Landau Spots Spleen Rep 2
Tutorial_Spatial_SC = Landau Spatial Cite Seq
Tutorial_SSC_MT = Stereo Cite Seq Mouse Thymus
Tutorial_Generated_data = Generated data

---

Code Explination:

Based on the code files you've provided for the GemVAE package and the tutorial notebook, I will outline the flow of the code and its components, which will then be represented in a mermaid flowchart.

Code Components and Flow Overview

1. Initialization (__init__.py):

   • Initializes the package by importing the main components: train_GEMVAE, various clustering methods, and utility functions like Cal_Spatial_Net, Stats_Spatial_Net, etc.

2. Utility Functions (utils.py):

   • Contains functions to calculate spatial networks, perform clustering using mclust from R, and plot weight values. It serves as support for preprocessing and analysis tasks.

3. Clustering (clustering.py):

   • Implements functions for spatial clustering, leveraging methods like mclust, leiden, and louvain for identifying cell clusters based on spatial and expression data.

4. Model Definition (model.py):

   • Defines the GATE model, which is a core part of GEMVAE, including encoder and decoder parts for genes and proteins, graph attention layers, and the variational autoencoder logic.

5. GEMVAE Core (GEMVAE.py):

   • Implements the GEMVAE class, orchestrating the model's training process, including setting up the placeholders, building the model, running epochs, and inferring results.

6. Training Script (Train_GEMVAE.py):

   • Provides a high-level interface to train the GEMVAE model on given data, handling data preparation, model initialization, training, and output processing.

7. Tutorial Notebook (Tutorial_Landau_SR1.ipynb):

   • A Jupyter notebook demonstrating how to use GemVAE for a specific dataset, likely guiding the user through data loading, model training, and analysis of results.