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

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Identifying long noncoding RNA (lncRNA) and microRNA (miRNA) interactions is vital for understanding gene regulation and diseases.
• Current graph neural network (GNN) methods face limitations in representation capacity, edge bias, and multimodal feature fusion.

Purpose of the Study:

• To develop an advanced heterogeneous graph attention model, HetGAT-LMI, for accurate prediction of lncRNA-miRNA interactions.
• To overcome limitations of existing GNNs by integrating diverse sequence and structural features.

Main Methods:

• Constructed a heterogeneous lncRNA-miRNA network with multi-level similarity and interaction edges.
• Fused multimodal RNA features (K-mer, G-gap, CTD, RNAfold structural) into unified representations.
• Employed GATv2 multihead attention and pairwise gated fusion for robust encoding and discrimination.

Main Results:

• HetGAT-LMI achieved high performance with AUC of 0.9585 and AUPR of 0.9467.
• SHAP analysis revealed key feature importance (CTD for miRNA, MFE for lncRNA).
• Case studies on HOTAIR and MALAT1 validated the model's biological relevance and external validity.

Conclusions:

• HetGAT-LMI offers enhanced accuracy, robustness, and interpretability in predicting lncRNA-miRNA interactions.
• The model serves as a valuable tool for high-throughput screening and generating hypotheses on interaction mechanisms.