CircGO: Predicting circRNA Functions Through Self-Supervised Learning of Heterogeneous Networks.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Circular RNAs (circRNAs) are crucial non-coding RNAs involved in various biological processes.
• Despite advances, the functional annotation of circRNAs remains a significant challenge.
• Automated methods are needed to explore the functions of newly discovered circRNAs.

Purpose of the Study:

• To develop a novel computational approach for predicting Gene Ontology (GO) functions of circRNAs.
• To leverage self-supervised learning on a circRNA-protein heterogeneous network for functional annotation.
• To improve the efficiency and accuracy of circRNA functional prediction.

Main Methods:

• Construction of a heterogeneous network integrating circRNA co-expression, circRNA-protein associations, and protein-protein interactions (PPI).
• Initialization of node features and pseudo-labels using graph processing techniques (walking, aggregation, clustering).
• Heterogeneous graph attention network for node feature learning and attention-based label propagation for pseudo-label refinement during self-supervised pre-training.

Main Results:

• The proposed method demonstrated superior performance in predicting circRNA GO terms compared to existing approaches on an independent test set.
• The study highlighted the critical role of network structure and initialization strategies in prediction accuracy.
• Analysis suggests potential improvements by integrating additional heterogeneous information and association networks.

Conclusions:

• The developed self-supervised learning framework effectively predicts circRNA GO functions.
• The findings emphasize the utility of heterogeneous network analysis and graph-based learning in functional genomics.
• The approach provides a valuable tool for exploring the functional landscape of circRNAs.