GCN-Based Heterogeneous Complex Feature Learning to Enhance Predictability for LncRNA-Disease Associations

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Computational models are crucial for predicting long non-coding RNA-disease associations (LDAs) to understand disease pathogenesis.
• Existing models struggle to capture complex features within biological networks.

Purpose of the Study:

• To propose HGCNLDA, a novel computational model for inferring LDAs.
• To enhance the extraction of essential features from biological network data.

Main Methods:

• Constructed a tripartite heterogeneous network (lncRNA-disease-miRNA network, LDMN).
• Employed a graph convolutional network (GCN)-based encoder for feature extraction.
• Utilized feature fusion with bipolymerization and attention mechanisms.
• Applied a bilinear-decoder for predicting association scores.

Main Results:

• HGCNLDA demonstrated superior performance over five existing models in 5-fold cross-validation on two datasets.
• Achieved high AUROC and AUPR values, particularly on a challenging dataset.
• Case studies confirmed HGCNLDA's practicality in identifying potential LDAs in cancer.

Conclusions:

• HGCNLDA effectively predicts lncRNA-disease associations by leveraging heterogeneous information fusion and GCNs.
• The model offers a practical tool for exploring disease pathogenesis and identifying potential therapeutic targets.
• The source code and data are publicly available for further research.