Predicting lncRNA-disease associations using multiple metapaths in hierarchical graph attention networks

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Long non-coding RNAs (lncRNAs) play crucial roles in regulating gene expression and are implicated in complex disease pathogenesis.
• Predicting lncRNA-disease associations is vital for understanding disease mechanisms but is challenging due to the vast number of lncRNAs and experimental limitations.
• Existing computational methods often overlook the information provided by intermediate nodes in network structures.

Purpose of the Study:

• To develop a novel deep learning model, MMHGAN, for predicting unknown lncRNA-disease associations.
• To leverage hierarchical graphical attention networks and multiple metapath types for enhanced feature extraction.
• To improve upon existing methods by considering a broader range of network information, including intermediate nodes.

Main Methods:

• Construction of a heterogeneous graph integrating lncRNA-disease-miRNA associations and homogeneous graphs for lncRNAs and diseases.
• Application of a multihead attention mechanism for aggregating features in homogeneous graphs.
• Selection and weighting of metapaths with different intermediate nodes in the heterogeneous graph to derive final embedded features.

Main Results:

• Achieved an average AUC of 96.07% and an average AUPR of 93.23% via fivefold cross-validation.
• Ablation experiments confirmed the significance of homogeneous graphs and varying intermediate node path weights.
• Case studies on lung cancer, esophageal carcinoma, and breast cancer showed high validation rates for predicted lncRNA-disease associations.

Conclusions:

• The MMHGAN model demonstrates superior performance compared to six existing prediction models.
• The model effectively predicts potential lncRNA-disease correlations, as validated by case studies.
• MMHGAN offers a promising computational approach for advancing lncRNA-disease association research.