Abstract
The interaction between circular RNAs (circRNAs) and RNA-binding proteins (RBPs) plays a crucial role in gene regulation; however, experimental identification is costly and inefficient. Current computational methods often overlook the structural features of circRNAs, thereby limiting prediction accuracy. To address these challenges, we propose GGCRB, a deep learning framework that integrates both sequence and structural features for predicting circRNA-RBP binding sites. Sequence features are captured through five encoding schemes (HFN, ND, NCP, DPCP, and Doc2Vec), followed by convolutional layers for local pattern extraction. Structural features are derived from base-pairing adjacency matrices generated by RNAstructure and modeled using graph convolutional networks and graph attention networks to learn topological dependencies. The fused representations are further processed by bidirectional LSTM and multihead attention modules to capture global interactions. Final predictions are made through pooling and softmax layers. Extensive experiments on 16 benchmark data sets demonstrate that GGCRB significantly outperforms existing models. Ablation studies and motif analyses further confirm its effectiveness, underscoring the importance of integrating structural and sequence information for accurate prediction of circRNA-RBP interactions.