Graph neural network for gene-disease link prediction

Abstract

This study explores the use of Graph Neural Networks Graph Neural Network (GNN) for identifying gene-disease associations through a Knowledge Graph derived from the Stanford Biomedical Network Dataset Collection, which includes originally over 21,000 connections. By representing genes and diseases as nodes, and their relationships as edges, we enable GNNs to uncover complex patterns and predict novel associations. This approach not only reveals previously unknown gene-disease links but also offers fresh insights into the molecular basis of diseases. The specific disease-gene association network analyzed comprises 7,813 nodes and 21,357 edges. Initial data preprocessing involved standardizing disease features and augmenting them using Large Language Models for attributes like disease class, alongside enhancing gene features by parsing chromosomal location data. The methodology evolved from heterogeneous graph models to a primary focus on homogeneous graph representations, unifying node types and their 512-dimensional features. Custom GNNs, employing backbones such as GraphConv, SAGEConv, and GATv2Conv, were developed, primarily utilizing a Multi-Layer Perceptron Multi-Layer Perceptron (MLP) based decoder for link prediction. This refined approach, particularly with simpler homogeneous GNN models, demonstrated significantly improved generalization and high validation performance, achieving an AUC-PR of 0.991 and an F1-Score of 0.950, effectively addressing earlier overfitting challenges. Evaluation employs comprehensive metrics including AUC-ROC, AUC-PR, F1-Score, and Precision@10. The research underscores the strong potential of tailored GNNs for discovering novel gene-disease links, with ongoing work focusing on further optimization and interpretability.