Hyperparameter tuning using bio-inspired algorithms in GCN and GAN models for link prediction in a PPI network

Overview

This project is made to compare different biologically inspired algorithms like Genetic Algorithm, Particle Swarm Optimization, Ant Colony Optimization, Artificial Bee Colony and classic ones like Grid Search, Simulated Annealing, Hill Climbing and Random Search as well as Optuma and Bayesian Optimization in hyperparameter optimization in Graph Convolutional Network and Generative Adversarial Network models for link prediction in a SNAPS protein-protein interaction network.

🧬 Protein-Protein Interaction Network: PP-Pathways

This repository explores the PP-Pathways dataset from the Stanford SNAP BioData collection. It represents a large-scale protein-protein interaction (PPI) network derived from pathway databases.

• Nodes: 21,554 proteins
• Edges: 342,338 interactions
• Data Type: Undirected, unweighted graph
• Source: Pathway-based protein associations
• Format: Edge list (.csv) with each row representing a protein-protein interaction

Graph visualization for the PPI network

This image was done using Cytoscape.

Results

The following results are obtained after 10 epochs of training in each GCN and GAN model.

Model	Algorithm	F1	AUC	Loss / Avg Loss	NDCG	Hidden Channels	Learning Rate	# Layers	Dropout	Time
GCN	None	0.8071	0.8775	1.2861	0.9875	256	0.01	2	0	1m 19s
	GA	0.8506	0.9125	1.2876	0.9913	73	0.0122	3	0.4	9m 20s
	PSO	0.8506	0.779	1.392	0.968	107	0.01339	3	0.55	11m 56s
	ABC	0.8496	0.8831	1.4161	0.9885	106	0.00691	3	0.15	14m 28s
	Simulated Annealing	0.8435	0.7841	1.3751	0.9685	160	0.00974	3	0.23	13m 18s
	Hill Climbing	0.843	0.9109	1.3705	0.9914	244	0.01102	3	0.66	9m 10s
	Random Search	0.8493	0.9147	1.2535	0.992	38	0.01378	3	0.18	12m 20s
	ACO	0.8419	0.9145	1.2698	0.9918	224	0.00215	3	0.7	4m 30s
	Bayesian Search	0.8504	0.9154	1.2674	0.9911	80	0.00785	4	0.1	10m 22s
	Grid Search	0.8497	0.9151	1.2031	0.992	64	0.01	3	0.3	9m 13s
	Optuna	0.8512	0.9166	1.2522	0.992	277	0.00153	4	0	22m 22s
GAN	None	0.7337	0.7528	-0.3044	0.9699	256	1e-4	-	0.3	1m 21s
	GA	0.7538	0.7772	0.0088	0.9723	459	0.0019	-	0.33	8m 59s
	PSO	0.7571	0.7781	0.0147	0.9706	512	0.002	-	0.31	15m 22s
	ABC	0.7545	0.7773	0.0421	0.9724	480	0.00136	-	0.22	8m 32s
	Simulated Annealing	0.7584	0.7583	-0.026	0.9699	412	0.002	-	0.1	18m 12s
	Hill Climbing	0.7559	0.7773	0.0015	0.9726	393	0.002	-	0.4	11m 47s
	Random Search	0.7541	0.7752	0.0432	0.9725	367	0.00166	-	0.2	8m 36s
	ACO	0.7790	0.7736	0.5720	0.9716	64	0.00167	-	0.4	13m 29s
	Bayesian Search	0.7505	0.7743	0.2257	0.9725	144	0.00062	-	0.35	9m 55s
	Grid Search	0.7428	0.763	0.0554	0.9712	512	0.0001	-	0	12m 6s
	Optuna	0.756	0.7817	0.2407	0.9738	373	0.00068	-	0.51	20m 24s

Quickstart

Prerequisites

• Python 3.8+
• scikit-learn
• PyTorch
• PyTorch Geometric
• NiaPy

Installation

git clone https://github.com/milagjurovska/PPI-link-prediction-with-optimized-gcn-and-gan.git
cd PPI-link-prediction-with-optimized-gcn-and-gan

The results can be displayed in the Jupyter Notebook provided, however if you want to run the code in Python only, there is a results.py file.