MFH-LPI: based on multi-view similarity networks fusion and hypergraph learning for long non-coding RNA-protein interactions prediction

Studies demonstrate that long non-coding RNAs (lncRNAs) and their protein interactions (LPIs) play crucial roles in regulating gene expression and participating in diverse biological processes. Aberrant expression of these interactions is closely associated with the initiation and progression of various diseases. Therefore, investigating LPI prediction is critical for elucidating disease mechanisms and identifying potential biomarkers and therapeutic targets. Given the high costs and limited efficiency of traditional biological methods, developing cost-effective and accurate computational models for LPI prediction becomes essential. Inspired by similarity network fusion and hypergraph learning, this study proposes a computational framework named MFH-LPI. First, we construct separate similarity networks for lncRNAs and proteins, then employ an attention mechanism to extract and fuse key features from these multi-view networks. Subsequently, we introduce a hypernode (randomly generated node) to establish a heterogeneous hypergraph integrating lncRNAs and proteins, thereby capturing richer node representations. Finally, we predict LPIs using a multilayer graph convolutional network (GCN) combined with a fully connected (FC) layer. We conduct several experiments on three datasets to validate the method's effectiveness. The experimental findings indicate that the suggested model is effective compared to existing processes and outperforms other approaches.