A Multisource Transformer-Guided Graph Representation Learning Framework for circRNA-Disease Association Prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Circular RNAs (circRNAs) exhibit stability and tissue-specific expression, making them promising disease biomarkers.
• Predicting circRNA-disease associations is vital for early diagnosis and treatment but challenged by complex data and traditional methods' limitations.

Purpose of the Study:

• To develop a high-accuracy computational model, MTGCDA, for predicting circRNA-disease associations.
• To address limitations in information integration, semantic expression, and information loss in existing prediction methods.

Main Methods:

• MTGCDA integrates multisource biological information into a heterogeneous graph with multiple node and edge types.
• Representation learning is performed using a heterogeneous graph neural network to capture latent semantic features.
• A multilayer heterogeneous graph convolutional network fuses node embeddings, followed by a CatBoost classifier for association scoring.

Main Results:

• MTGCDA achieved an Area Under the Curve (AUC) of 0.9756 on the CircR2Disease dataset, outperforming existing methods.
• 17 out of 20 predicted circRNA-disease associations were validated by literature reports, confirming the model's accuracy and practicality.

Conclusions:

• The MTGCDA model demonstrates superior performance in predicting circRNA-disease associations.
• MTGCDA offers a practical and accurate computational approach for identifying potential circRNA-disease links, supporting biomarker discovery and therapeutic strategies.