HetGAT-LMI: Interpretable Heterogeneous Graph Attention Method for Predicting lncRNA–miRNA Interactions

Identifying long noncoding RNA (lncRNA) and microRNA (miRNA) interactions is crucial for understanding post-transcriptional regulation and disease mechanisms. Existing graph neural network (GNN) based methods are often constrained by the limited representational capacity of single similarity graphs, noise-induced edge bias, and inadequate fusion of sequence and structural information. To address these limitations, we propose HetGAT-LMI, a heterogeneous graph attention model. At the graph level, we construct a lncRNA-miRNA heterogeneous network comprising lncRNA-lncRNA and miRNA-miRNA similarity edges alongside lncRNA-miRNA interaction edges. At the representation level, we fuse K-mer, G-gap, CTD, and structural features from RNAfold, into unified multimodal representations. At the encoding level, we apply GATv2 multihead attention to each relation type and employ parameter-free mean aggregation for robust fusion. At the discrimination level, we introduce a pairwise gated fusion mechanism to adaptively emphasize interaction-relevant channels, culminating in probability outputs via an MLP. Balanced sampling and early stopping strategies are incorporated to mitigate class imbalance and overfitting. HetGAT-LMI achieves AUC of 0.9585 and AUPR of 0.9467, demonstrating superior overall performance compared to several representative methods. SHAP interpretability analysis reveals that CTD features dominate on miRNA side, while MFE features are most influential on lncRNA side, consistent with established biological knowledge. Case studies on HOTAIR and MALAT1 further validate the external validity of our model. These results indicate that HetGAT-LMI synergistically enhances accuracy, robustness, and interpretability, providing a reliable tool for high-throughput screening of candidate molecules and hypothesis generation regarding interaction mechanisms.