Abstract
Accurate identification of RNA-protein binding sites is crucial for understanding gene regulation and disease mechanisms. However, existing deep learning models still face challenges in synergistically modeling multi-scale sequence motifs and dynamic structures. To address this, we propose GMC-Bind, a multimodal framework combining multi-scale window convolution, graph attention, and bidirectional cross-attention mechanisms, which achieves efficient synergistic modeling of local sequence features and global structural dependencies. The framework employs multi-scale convolution to extract variable-length motif features, utilizes graph attention to encode the spatial topology of RNA secondary structures, and facilitates cross-modal interaction between sequence and structural features through a bidirectional cross-attention mechanism based on positional encoding, ultimately generating binding site predictions. Extensive experimental results on the RBP-24 dataset demonstrate that GMCBind achieves optimal performance in 17 out of 24 datasets, with an average AUC of 95.8%, significantly outperforming existing baseline methods and providing an effective new approach for RNA-protein binding site identification. The source code of our method can be found in https://github.com/benqin66/GMC-Bind/.
Funding Information
  • Adult Continuing Education Scientific Research Planning Project of China Adult Education Association (2025-0121Y)
  • 2025 Special Open Project of Applied Journals (CACJ202508ZZ)
  • High-Level Talent Research Startup Foundation Project of Wuhan Software Engineering Vocational College (KYQDJF2026006)

This publication has 32 references indexed in Scilit: