A Multi-Layer Deep Graph Attention Model for Decoding lncRNA–miRNA–mRNA Communication Networks

Abstract

Dysregulation of lncRNA–miRNA–mRNA (LncM2) communication networks, a key mechanism of competing endogenous RNA (ceRNA) crosstalk, plays a critical role in cancer progression. Current computational approaches are limited by their dependence on predicted interactions, lack of interpretability, and inability to effectively address transcriptomic heterogeneity. To overcome these challenges, we present MDGAM-LncM2Net, a Multi-layer Deep Graph Attention Model for reconstructing directed, context-specific LncM2 communication networks from multi-omics data. Our framework combines experimentally validated RNA interactions with transcriptomic profiles through a comprehensive analytical pipeline incorporating differential expression analysis, hypergeometric testing, Spearman correlation, and Fast Causal Inference (FCI) for directional relationship inference. MDGAM-LncM2Net leverages Graph Attention Networks (GATs) to capture topological features and Bidirectional LSTMs to model temporal expression patterns, enabling interpretable survival prediction. When applied to four cancer types (LSCC, OvCa, PrCa, and GBM), the model identified 14 pan-cancer driver lncRNAs and 11 conserved prognostic gene clusters with cancer-specific communication nodes strongly associated with patient survival. The proposed framework achieves an F1-score of 90.29% and AUC of 90.60%, representing an average improvement over the state-of-art methods. MDGAM-LncM2Net provides a robust platform for elucidating LncM2 network, with significant potential for advancing precision oncology through biomarker discovery and targeted therapy development.