A Hybrid Graph Convolutional Network and XGBoost Framework for miRNA–Disease Association Prediction

Abstract

Predicting potential associations between microRNAs (miRNAs) and diseases is essential for deciphering complex pathogenic mechanisms and advancing personalized medicine. However, many existing computational methods fail to adequately capture high-order topological structures within heterogeneous biological networks, limiting their predictive performance. In this study, we propose GCN-XGB, a novel hybrid computational framework that integrates a two-layer Graph Convolutional Network (GCN) with Extreme Gradient Boosting (XGBoost) to improve the accuracy of miRNA-disease association prediction. Specifically, we first input k-mer derived initial features of miRNAs, BioBERT-derived initial features of diseases, and the computed miRNA-miRNA similarity, disease-disease similarity, and miRNA-disease association information into the two-layer GCN. Through deep feature propagation and aggregation of second-order neighbor information, the GCN integrates these inputs to generate enhanced discriminative high-dimensional embeddings, providing more informative and discriminative features for the downstream XGBoost classifier. Experimental results under rigorous 10-fold cross-validation demonstrate that GCN-XGB consistently outperforms several state-of-the-art baseline models when all models are trained on the same GCN-enhanced features, achieving a superior AUC of 0.9601 and AUPR of 0.9587. Furthermore, case studies on prevalent neoplasms (e.g., Breast, Colon, and Lung Cancer) confirm the framework's efficacy in discovering novel associations even in the absence of prior clinical data. Our findings suggest that GCN-XGB is a powerful and reliable tool for identifying potential disease-related miRNAs and prioritizing candidates for experimental validation.