GSASVM-RBPs: Predicting miRNA-binding protein sites with aggregated multigraph neural networks and an SVM

RNA-binding proteins (RBPs) are a class of proteins with RNA-binding domains involved in regulating various cellular processes, such as RNA processing, transport, splicing, translation, and stability. They play crucial roles in maintaining cell differentiation, development, apoptosis, and inflammation. Although some deep learning graph neural network algorithms achieve high accuracy in RBP prediction, most methods do not fully consider the characteristics of heterogeneous graphs, often resulting in information loss during the encoding and decoding processes. We aimed to develop a graph model that can aggregate heterogeneous information on miRNAs and proteins and efficiently predict RBPs. In this study, we introduce a method that combines GraphSAGE and graph attention network techniques for encoding and aggregating relevant graph information, and uses an SVM for decoding predictions, which we refer to as the GSASVM. We collected data, constructed RBP datasets for humans and mice, and compared our method against seven state-of-the-art methods on these datasets. Our approach demonstrated superiority across various evaluation metrics, achieving AUC and PRC values of 98.46% and 97.98%, respectively, on the human dataset and 97.38% and 97.43%, respectively, on the mouse dataset. Additionally, we conducted two specific prediction studies on human proteins, providing some RBP results through case analyses. These experiments validate the potential of this method as a novel research tool for RBP-related tasks. The model optimizes the aggregation of encoded representations and effectively utilizes complex graph structures related to RBPs for feature extraction and decoding. The experimental results also verify that our new framework can effectively predict RBP binding sites, potentially facilitating further downstream analysis in biomedical and biotechnology applications.