Predict Phyiscal Protein Interactions using Graph Neural Networks

This repository contains code for predicting physical interaction of proteins using direct coupling analysis and graph neural networks.

Setup:

This code was designed to run on the s3it cluster to avoid out of memory (OOM) issues.

GPU Usage:

Because of the large datasets involved data-loading and training can take a large amount of time. To run on s3it GPU cluster, make sure to have a GPU compatible version of tensorflow installed before running the code. Please visit s3it-GPU instructions.

Installation:

This requires a valid installation of Anaconda or miniconda. Create the Python environment as described below:

cd configs
conda env create -f env.yml 
conda activate gcn_env

Data:

You can find the data paths within the gcn_generator.py file (see scripts directory).

Data Preparation:

To generate the inter protein graphs, run:

cd src/scripts
bash run_graph_generator.sh

Running the Model:

To run the model simply run:

cd src/scripts
bash run_graph_prediction.sh

Example Results:

Running the GCN on the E.coli dataset shows a learnable signal and good predictive performance

Model:

Models	Resources
Neural Net (Spektral GCN)	Spektral model

Potential Conflicts:

There may be issues running this code from a local machine. It was designed to run on the s3it cluster.

Major Dependencies:

The conda environemnt provided should contain all of these requirements. If not, you can find them at the following sources.

Dependency	Installation
Spektral (cpu, linux)	Pypi