GMC-Bind: A Multimodal Framework for RNA-Protein Binding Site Prediction With Bidirectional Cross-Attentional Fusion

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate RNA-protein binding site identification is vital for understanding gene regulation and disease.
• Current deep learning models struggle to effectively combine multi-scale sequence motifs and RNA dynamic structures.

Purpose of the Study:

• To develop an advanced multimodal deep learning framework, GMC-Bind, for enhanced RNA-protein binding site identification.
• To synergistically model both local sequence features and global structural dependencies in RNA.

Main Methods:

• GMC-Bind employs multi-scale window convolution for variable-length motif extraction.
• Graph attention networks encode the spatial topology of RNA secondary structures.
• Bidirectional cross-attention mechanisms with positional encoding integrate sequence and structural information.

Main Results:

• GMC-Bind achieved optimal performance on 17 out of 24 datasets in the RBP-24 benchmark.
• The framework demonstrated a high average AUC of 95.8%.
• GMC-Bind significantly outperformed existing baseline methods for RNA-protein binding site identification.

Conclusions:

• GMC-Bind offers an effective and novel approach for identifying RNA-protein binding sites.
• The synergistic modeling of sequence and structure advances the field of RNA-protein interaction analysis.
• This method holds potential for applications in understanding gene regulation and disease mechanisms.