MTGCDA: Enabling Accurate CircRNA-Disease Association Prediction Through Transformer-Guided Multi-Source Graph Learning

Circular RNA (circRNA) has a stable structure and tissue-specific expression, which is of great value in the diagnosis and treatment of diseases. However, complex biological relationships and heterogeneous data hinder the prediction of circRNA-disease associations, resulting in challenges such as weak semantic representation and information loss. To address this problem, we propose a transformer-based multi-source heterogeneous graph model MTGCDA. The model first aggregates various biological data sources related to circRNA and disease to construct a heterogeneous graph containing various types of nodes and relationships. By applying a specialized heterogeneous graph neural network, the unique structural and contextual properties of different biological entities are captured. Subsequently, the circular RNA and disease node embeddings derived from multi-layer heterogeneous graph convolutional networks are combined to form a comprehensive joint representation. The fused embeddings are then processed by a CatBoost classifier to accurately estimate the likelihood of potential associations. Experiments on the CircR2Disease dataset show that MTGCDA achieves an AUC of 0.9756, outperforming existing methods. In addition, 9 of the 10 best predictions have been validated by literature, demonstrating the effectiveness and biological relevance of the model.