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

Circular RNAs (circRNAs) possess structural stability and tissue-specific expression patterns, making them potential disease biomarkers. Exploring the associations between circRNAs and diseases is crucial for early diagnosis and targeted treatment. However, due to the complexity of biological relationships and the presence of multisource heterogeneous data, traditional prediction methods often face challenges such as insufficient information integration, limited semantic expression, and information loss. To address these challenges, this paper proposes a computational model, MTGCDA, based on a multisource heterogeneous graph transformer for high-accuracy prediction of circRNA-disease associations. Specifically, MTGCDA first integrates multisource biological information about circRNAs and diseases, constructing a heterogeneous graph containing multiple node types and multiple edge relationships. Representation learning of the graph structure is performed using a heterogeneous graph neural network, fully exploiting the latent semantic features of different node types. The model then fuses the embeddings of circRNA and disease nodes in a multilayer heterogeneous graph convolutional network to construct a joint feature representation, which is then fed into a CatBoost classifier to accurately score circRNA-disease associations. Experimental results on the CircR2Disease data set demonstrate that the MTGCDA model achieves an AUC of 0.9756, significantly outperforming several existing methods. At the same time, among the 20 selected circRNA-disease association pairs, 17 pairs were verified by literature reports, further demonstrating the high accuracy of the model in predicting potential associations and its biological practicality.