Stacked Graph Attention Network With Temporal Modeling for lncRNA-miRNA Association Network

Long non-coding RNAs (lncRNAs) and microRNAs (miRNAs) play critical roles in regulating biological processes and disease mechanisms through their interactions, known as lncRNA-miRNA associations (LMAs). Accurate prediction of LMAs is essential for understanding disease progression, yet existing computational methods often exhibit inadequate feature integration, limited generalizability, and insufficient modeling of complex interactions. This study introduces the Stacked Graph Attention Network with Temporal Modeling (SGAT-TM), a novel framework that integrates statistical, graph-structural, and sequence-derived features of lncRNAs and miRNAs. SGAT-TM utilizes a multilayer Graph Attention Network (GAT), enhanced by self-attention, a multilayer perceptron (MLP), and a Gated Recurrent Unit (GRU) to capture LMA patterns, preserve long-range dependencies, and produce robust representations. Evaluation on benchmark datasets demonstrates that SGAT-TM outperforms state-of-the-art methods in predictive accuracy, robustness, and generalizability. SGAT-TM enhances LMA network analysis and provides significant insights into lncRNA-miRNA associations in biological and disease contexts.