Multi-view graph neural network with cascaded attention for lncRNA-miRNA interaction prediction

Abstract

Identifying interactions between long non-coding RNAs (lncRNAs) and microRNAs (miRNAs) reveals the mechanisms of biological processes, thereby contributing to disease diagnosis and treatment. Recently, graph neural networks (GNNs) have achieved remarkable progress in this task due to their consideration of both node attributes and graph topology. Nevertheless, existing GNN-based methods use only one type of node attribute, and the possible bias of a single view leads them to learn suboptimal node representations. Moreover, the underlying mechanisms of action between lncRNAs and miRNAs are complex. Ignoring the importance of neighboring nodes to the target node and the influence of different order neighborhood information makes them fail to learn satisfactory topological information. To this end, we propose a novel Multi-view Graph Neural Network with Cascaded ATtention (MGCAT) for lncRNA-miRNA interaction (LMI) prediction, where cascaded attention is a key ingredient consisting of view-level, node-level, and layer-level attentions. Specifically, we first construct a multi-attributed LMI graph to fully characterize lncRNAs and miRNAs, where nodes have multiple node attributes (i.e., multi-view features). Next, view-level attention dynamically integrates multi-view features to capture the inherent attribute information of nodes. Then, node-level attention iteratively aggregates the neighborhood information of each node. Finally, layer-level attention adaptively combines integrated features and different order neighborhood information to obtain informative node representations. Extensive experiments on four benchmark datasets show that MGCAT consistently outperforms recent state-of-the-art methods. Further case studies demonstrate the potential ability of MGCAT to identify novel LMIs. Code and datasets are publicly available at https://github.com/ai4slab/mgcat.