Predicting MiRNA-MRNA Interactions via Multi-Scale Feature Integration and Dual-Layer Graph Attention Networks

The prediction of miRNA-mRNA interactions is fundamental to elucidating gene regulatory mechanisms and disease pathogenesis. This study proposes GMLA, a computational framework for this predictive task. The architecture first utilizes an autoencoder to derive compressed, low-dimensional feature embeddings for miRNAs and mRNAs. These embeddings then populate a heterogeneous graph, where a dual-layer Graph Attention Network (DLGAT), augmented with residual connections, is employed to capture intricate topological dependencies. A Jumping Knowledge Network (JKNet) that leverages a multi-head self-attention mechanism aggregates these layer-specific representations, enhancing the expressive capacity of the model. The efficacy of the model was systematically evaluated across several benchmark datasets characterized by diverse scales and distributions. On the principal MTIS-10317 dataset, GMLA yielded an AUC of 0.8867 and an AUPR of 0.8865. Ablation studies and parameter sensitivity analyses validated the functional contribution of each architectural component and delineated optimal hyperparameter configurations. Case studies involving the PTEN gene and miR-21-5p were conducted to evaluate the utility of the model in a practical setting, substantiating its capacity to identify biologically relevant interactions.