PPI-predictor

A graph neural network (GNN) model to predict protein-protein interactions (PPI) with no sample features

Overview

This is a simple graph convolutional network (GCN) to predict the protein-protein interactions. There are 2 datasets: a large one and a small one. The useful informations in the dataset are only known protein-protein interactions and the bioinformatic database query of proteins. Since crawling more informations from the database is troublesome, in this project, I want to predict PPIs with only there known interaction relationships, so GCN is utilized.

Software requirements

• Python 3.8.3
• PyTorch 1.6.0
• torch_geometric

Repo content explanation

• dataset folder contains two .txt files: a larger dataset and a small dataset.
• train.py is the script defining the GCN model and training it.
• metrics.py: compute the metrics for performance evaluation.
• topInteract.py: choose the protein-protein pairs with highest score as the predicted PPIs.
• *_files folders contain the output files when training based on different dataset.