Adaptive Meta-Path Selection Based Heterogeneous Spatial Enhancement for circRNA-Disease Associations Prediction

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Circular RNAs (circRNAs) are key regulators of gene expression and disease via miRNA sponging.
• Current methods for analyzing circRNA interactions rely on manual meta-path construction, often introducing noise and hindering feature discovery.
• Improving feature distribution and representation is crucial for robust algorithmic generalization in biological data analysis.

Purpose of the Study:

• To develop an adaptive meta-path selection method (AdaMH) for enhanced circRNA analysis.
• To address challenges in manual meta-path selection, data sparsity, and feature representation in biological networks.
• To improve the predictive accuracy and generalization ability of models analyzing circRNA functions and disease associations.

Main Methods:

• Proposed AdaMH, an adaptive meta-path selection strategy that automatically identifies relevant heterogeneous paths during training.
• Incorporated graph contrastive learning to introduce controlled noise, ensuring even feature distribution in sparse data.
• Utilized a multi-head attention mechanism to capture complex relationships within high-dimensional heterogeneous feature spaces.

Main Results:

• AdaMH demonstrated superior performance compared to state-of-the-art algorithms.
• The method achieved a performance threshold exceeding 0.95 across seven distinct evaluation metrics.
• AdaMH effectively identified highly correlated hidden features and improved representation capability.

Conclusions:

• AdaMH offers a significant advancement in analyzing circRNA-mediated gene regulation and disease associations.
• The adaptive approach and noise injection improve model robustness and predictive power.
• This method provides a more effective way to capture complex biological interactions from heterogeneous network data.