GSSCMI: Efficient Co-Attention and Multimodal Contrastive Learning for Enhanced circRNA-miRNA Interaction Prediction

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Current circRNA-miRNA interaction prediction methods often neglect spatial folding information and struggle with multimodal consistency and discrimination.
• Traditional attention mechanisms are computationally expensive and lack balanced feature fusion, limiting prediction performance.

Purpose of the Study:

• To propose GSSCMI, a novel method that leverages spatial folding, sequence, and similarity features for improved circRNA-miRNA interaction prediction.
• To enhance intra-modal consistency and inter-modal discrimination using multimodal contrastive learning.
• To introduce an efficient co-attention mechanism for balanced and computationally inexpensive feature fusion.

Main Methods:

• Information integration of similarity, sequence, and secondary structure features.
• Multimodal contrastive learning for intra-modal consistency and inter-modal discrimination.
• An efficient co-attention mechanism for unified and fine-grained attention-based feature fusion.

Main Results:

• GSSCMI achieved superior performance over existing methods, with notable improvements in MCC (9.53%) and F1 (6.40%).
• Ablation studies confirmed the efficiency of the co-attention mechanism, reducing convergence iterations by ~58% and computational cost by 42.94% while enhancing MCC by 13.32% and ACC by 6.69%.
• Identified regulatory sites via structure-based A-to-I editing and visualized sequence-level dependencies.

Conclusions:

• GSSCMI offers a significant advancement in circRNA-miRNA interaction prediction by effectively integrating multimodal data and employing an efficient co-attention mechanism.
• The proposed method demonstrates improved accuracy, efficiency, and provides insights into regulatory mechanisms.
• GSSCMI represents a promising tool for understanding circRNA-mediated gene regulation.