GNN-Disease: Learning Disease Similarity from Gene Graph Databases

This is the course project repo for CSCI-5120 Course in The Chinese University of Hong Kong. The project title is: GNN-Disease: Learning Disease Similarity from Gene Graph Databases

Group Name: AIH

Group Member: Yixuan Wang, Yimin Fan, Liang Hong, Jiyue Jiang

Project Description: In this project, Our goal is to quantify the similarity between disease using existing graph databases. First, we collect data from two common human gene databases called HumanNet and Go Database. Then we use Graph neural network to embed th gene graph and perform contrastive learning in the gene embedding space to obtain high-quality gene embeddings. In the inference stage, we first collect the genes related to specific disease, and use average pooling to get the disease embedding and analyze the similarity across disease through the similarity of disease embeddings

Environment Setup

conda create -n gnn_disease python=3.9
pip install -r requirements.txt

Usage

Dataset Preparation Download the dataset from this url and unzip it, put it into the data folder
```
https://drive.google.com/file/d/1vnWfWbR_lb6uZ-mC0m7RXZ8dPlB-o5kR/view?usp=sharing
```

Model Training Our model supports multiple graph neural network. The edges in the HumanNet graph database do not contain much information, so we can use the below networks for encoding

GCN https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.conv.GCNConv.html#torch_geometric.nn.conv.GCNConv
GAT https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.conv.GATConv.html#torch_geometric.nn.conv.GATConv
GATv2 https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.conv.GATv2Conv.html#torch_geometric.nn.conv.GATv2Conv
EG https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.conv.EGConv.html#torch_geometric.nn.conv.EGConv
GraphGEN https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.conv.GENConv.html#torch_geometric.nn.conv.GENConv

The edges in the GO database contain information about the relationship between genes, so we can use the below relation aware network for encoding

RGCN https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.conv.GCNConv.html#torch_geometric.nn.conv.GCNConv
FastRGCN https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.conv.FastRGCNConv.html#torch_geometric.nn.conv.FastRGCNConv
RGATConvNet  https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.conv.RGATConv.html#torch_geometric.nn.conv.RGATConv

To run our code using the above mentioned network, we can just do the following

bash submit.sh GCN RGCN

You can replace GCN with (GAT,GATv2,EG,GraphGEN) and replace RGCN with (FastRGCN, RGATConvNet)

Visualization After running the model, we can visualize the loss curve and the AUROC, AUPRC curve using the following code. Assume that you have already run for all combinations of HumanNet encoder and GO database encoder
```
 python visualization/plot.py --input_dir runs  --output_dir visualization \
     --plot_auprc True --plot_loss True --plot_auroc True
```
if you want to indiviually plot the result curve for one specific combination, you can run the follwowing code
```
 python visualization/plot_individual.py --g_encoder GCN \
    --kg_encoder RGCN --input_dir runs  --output_dir visualization \
     --plot_auprc True --plot_loss True --plot_auroc True
```

We alos provide a jupyter notebook notebook.ipynb for you to explore the pipeline of this project. We provide format_code.sh for formatting the code in this repo.

Acknowledgement

This codebase is built upon the project https://github.com/yhchen1123/CoGO

fym0503/GNN-disease

GNN-Disease: Learning Disease Similarity from Gene Graph Databases

Environment Setup

Usage

Acknowledgement