CircSSNN: circRNA-binding site prediction via sequence self-attention neural networks with pre-normalization

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Circular RNAs (circRNAs) function as transcription templates, making their interactions with RNA-binding proteins (RBPs) crucial for disease mechanisms.
• Current circRNA-RBP identification models, often based on CNNs or RNNs, struggle with parallelism, stability, and capturing long-range dependencies in RNA sequences.

Purpose of the Study:

• To develop an advanced computational model for accurate circRNA-RBP identification.
• To overcome the limitations of existing methods in analyzing deep semantic features of RNA sequences.

Main Methods:

• A novel method utilizing a self-attention mechanism to extract deep semantic features from RNA sequences.
• Construction of the CircSSNN model, integrating circRNA sequence representations with statistical distributions, static local contexts, and dynamic global contexts.
• A stable and efficient network architecture designed to capture long-term dependencies by reducing sequence position distances.

Main Results:

• Experiments on 37 circRNA datasets demonstrated superior stability, parallelism, and prediction performance of CircSSNN compared to existing models.
• The CircSSNN model showed effective transferability to lncRNA datasets without task-specific fine-tuning, highlighting its adaptability.
• The model achieved favorable results, indicating its robustness and efficiency.

Conclusions:

• CircSSNN is a powerful and scalable tool for circRNA-RBP identification with high performance.
• The model's ability to perform without task-oriented fine-tuning simplifies bioinformatics analysis and lowers the threshold for hyperparameter tuning.
• CircSSNN offers a promising solution for advancing research in circRNA-RBP interactions and related disease studies.