Predicting lncRNA-Disease Associations Based on a Dual-Path Feature Extraction Network with Multiple Sources of...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Identifying long noncoding RNA-disease associations (LDAs) is vital for disease management.
• Wet experiments for LDA discovery are costly and time-consuming.
• Computational methods offer an efficient alternative for predicting LDAs.

Purpose of the Study:

• To introduce DPFELDA, a dual-path feature extraction network for accurate LDA prediction.
• To address node feature oversmoothing in graph neural network-based LDA prediction.
• To integrate multi-source information for improved LDA prediction.

Main Methods:

• Constructed a dual-view lncRNA-disease structure and a heterogeneous lncRNA-disease-miRNA network.
• Employed a dual-path network combining GCN, CBAM, and node aggregation for topology feature extraction.
• Utilized a Transformer model for node-specific features and XGBoost for final LDA prediction.

Main Results:

• DPFELDA significantly outperformed existing benchmark models on multiple datasets.
• The model effectively alleviated node feature oversmoothing issues common in graph learning.
• Ablation studies validated the contribution of individual modules, and case studies confirmed predictive accuracy.

Conclusions:

• DPFELDA provides a robust and accurate computational approach for predicting lncRNA-disease associations.
• The method enhances understanding of disease mechanisms by identifying novel lncRNA-disease links.
• DPFELDA represents a significant advancement in computational bioinformatics for disease research.