HybridGNN: a graph neural network approach for human miRNA-disease association prediction.

Abstract

MicroRNAs (miRNAs) are small non-coding RNAs, typically 18-24 nucleotides in length, that play a pivotal role in RNA silencing and the post-transcriptional regulation of gene expression by targeting messenger RNAs (mRNAs). Dysregulation of these miRNAs has consistently been implicated in the onset and progression of a variety of complex human diseases. In this study, we propose a novel HybridGNN model that integrates a Graph Convolutional Network (GCN), a Graph Attention Network (GAT), and Matrix Decomposition with Matrix Factorization (MDMF) to predict potential miRNA-disease associations (MDAs). We incorporate five types of similarity in which three are derived from miRNAs and two are derived from diseases, to comprehensively explore and optimize multi-source feature information. The complementary interactions among these modules also help to mitigate the oversmoothing problem. The model utilizes neighboring nodes in a heterogeneous network to generate node embeddings via a message-passing mechanism. To improve computational efficiency, we employ a mini-batch gradient descent approach that partitions the graph into smaller sub-graphs, thereby enhancing the model's accuracy, speed, and scalability. As a result of these advanced techniques, HybridGNN achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.9715 using a dot-product classifier, outperforming several existing methods and underscoring its potential as a robust and accurate tool for predicting MDAs. Code and data are freely available at https://github.com/mbasharatahmad/HybridGNN-miRNA-disease/.