GNN-XG: Essential Protein Prediction Based on Graph Neural Networks and XGBoost

They are essential proteins that regulate cellular processes and maintain physiological functions. A successful drug target discovery and investigation of disease mechanisms depend on the accuracy of their prediction. Protein-protein interaction (PPI) networks have emerged rapidly in recent years, allowing structural prediction of essential proteins to be conducted. However, PPI network data often contain false positives and false negatives edges. Effectively and thoroughly mining the topological information of PPI networks has become critical to improving the performance of essential protein prediction algorithms. This study proposes GNN-XG, a key protein prediction method that integrates Graph Neural Networks (GNN) and XGBoost. First, the edges of the PPI network are weighted using gene expression data, thereby establishing the graph structure. Then, a weighted graph convolutional network (Weighted GCN) is employed to extract low-dimensional node embeddings. Subsequently, the Synthetic Minority Oversampling Technique (SMOTE) is applied to address class imbalance. Finally, the processed node embeddings are fed into an XGBoost classifier to identify key proteins. The experimental results indicate that the proposed framework consistently surpasses baseline approaches on various evaluation metrics, thereby validating its capability in essential protein prediction.