Predicting CircRNA-Disease Associations via Feature Convolution Learning With Heterogeneous Graph Attention Network.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Understanding circular RNA (circRNA) roles in disease pathogenesis is crucial.
• Experimental identification of circRNA-disease associations (CDAs) is resource-intensive.
• Existing computational prediction methods for CDAs have limitations.

Purpose of the Study:

• To develop a novel computational framework, GATCL2CD, for predicting unknown circRNA-disease associations (CDAs).
• To enhance the accuracy and efficiency of CDA prediction compared to existing methods.

Main Methods:

• Constructed a heterogeneous graph integrating Gaussian interactive profile kernel (GIP) similarity, semantic similarity, circRNA sequence similarity, and function similarity.
• Employed a feature convolution learning framework with a multi-head dynamic attention mechanism for node feature aggregation.
• Utilized a single-layer convolutional neural network (CNN) for extracting higher-order features.
• Implemented a pairwise element-wise product and a multilayer perceptron (MLP) for CDA inference.

Main Results:

• GATCL2CD demonstrated superior performance over five state-of-the-art methods in 5-fold cross-validation across three datasets.
• Case studies validated GATCL2CD's effectiveness in identifying potential disease-related circRNAs.

Conclusions:

• GATCL2CD offers a powerful and efficient computational approach for predicting circRNA-disease associations.
• The framework aids in uncovering disease mechanisms and identifying novel therapeutic targets through circRNA research.