Integrating Transformer and Graph Attention Network for circRNA-miRNA Interaction Prediction.

Area of Science:

• Computational biology
• Molecular biology
• Bioinformatics

Background:

• CircRNA-miRNA interactions (CMI) are vital in cellular gene regulation.
• Aberrant CMI is linked to various diseases.
• Existing prediction models overlook molecular attributes and global network structures.

Purpose of the Study:

• To develop an advanced computational model for predicting CMI.
• To overcome the limitations of existing methods by incorporating multi-feature fusion.

Main Methods:

• Proposed EGATCMI, a model combining transformer and graph attention networks.
• Utilized Word2vec for pre-training circRNA and miRNA sequences to capture feature representations and similarity.
• Employed a self-attention mechanism to extract global structural features from the CMI network.

Main Results:

• EGATCMI achieved high prediction accuracy, with AUC values of 0.9106 and 0.9470 on benchmark datasets.
• The model outperformed existing methods in predicting CMI.
• Case studies showed 80% accuracy in predicting disease-related miRNA-circRNA interactions.

Conclusions:

• EGATCMI effectively integrates multi-modal features for accurate CMI prediction.
• The model demonstrates significant potential as a tool for biological research and candidate screening.