Predicting miRNA-Disease Associations Based on Spectral Graph Transformer With Dynamic Attention and Regularization

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) are critical in human disease pathogenesis.
• Graph neural networks (GNNs) show promise for analyzing miRNA-disease relationships.
• Existing GNN methods struggle with effectiveness and node sensitivity.

Purpose of the Study:

• To develop an advanced deep learning model for accurate miRNA-disease association (MDA) prediction.
• To overcome limitations of current GNN-based MDA prediction methods.
• To introduce DARSFormer, a model leveraging dynamic attention and spectral graph Transformers.

Main Methods:

• Constructed a miRNA-disease heterogeneous network.
• Applied spectral decomposition and orthogonal GNN for feature refinement.
• Utilized a graph Transformer with dynamic attention for feature amalgamation.
• Employed a projection layer to calculate MDA scores.

Main Results:

• DARSFormer achieved an AUC of 94.18% on HMDD v2.0 and 95.27% on HMDD v3.2.
• Demonstrated high accuracy in predicting top associated miRNAs for colorectal, esophageal, and prostate cancers.
• Validated predictions against external databases (dbDEMC, miR2Disease).

Conclusions:

• DARSFormer significantly improves the accuracy of miRNA-disease association prediction.
• The model's dynamic attention and spectral graph Transformer components are key to its performance.
• DARSFormer offers a robust tool for advancing research in disease-related miRNA analysis.