RiGAT-CMI: RiGAT-CMI: Enhanced Prediction of Circular RNA–MicroRNA Interactions via Pre-trained RNA Language Models and Graph Attention Networks

The complex regulatory network of circular RNA (circRNA) and microRNA (miRNA) interactions (CMI) plays a pivotal role in gene expression and disease pathogenesis. While computational models have emerged as efficient alternatives to labor-intensive wet-lab experiments, existing methods often rely on static sequence encodings and struggle to capture the high-order collaborative dependencies inherent in CMI networks. In this study, we propose RiGAT-CMI, a novel deep learning framework that integrates the large-scale pre-trained RNA language model, RiNALMo, with Graph Attention Networks (GAT) for accurate CMI prediction. By leveraging the contextual embeddings from RiNALMo-giga, our model captures intricate biological features beyond simple sequence motifs. These representations are then processed through a multi-layer GAT architecture with residual connections, explicitly mining high-order relational signals in the bipartite interaction graph. Comprehensive experiments on benchmark datasets demonstrate that RiGAT-CMI achieves state-of-the-art performance. On the large-scale CMI-20208 dataset, our model yielded an impressive AUC of 0.9583, an AUPR of 0.9458, and an MCC of 0.8153 on an independent test set. Comparative analysis reveals that the RiNALMo-giga scale significantly outperforms smaller variants, providing more robust features for downstream graph learning. Furthermore, ablation studies investigating the number of GAT layers and alternative architectures, such as the Kolmogorov-Arnold Network (KAN), highlight the stability and generalization capability of our proposed 3-layer GAT configuration in mitigating overfitting. Our findings suggest that RiGAT-CMI is a powerful tool for discovering novel circRNA-miRNA associations, offering valuable insights into intracellular regulatory mechanisms.