Nucleotide-level circRNA-RBP binding sites prediction based on hybrid encoding scheme and enhanced feature extraction

Area of Science:

• Biochemistry
• Genomics
• Computational Biology

Background:

• Circular RNAs (circRNAs) regulate gene expression and cell functions through interactions with RNA binding proteins (RBPs).
• Accurate prediction of circRNA-RBP binding sites is vital for understanding disease mechanisms.
• Existing methods often lack nucleotide-resolution accuracy and fail to capture positional or feature interaction information.

Purpose of the Study:

• To develop a deep learning framework, circdpb, for precise prediction of circRNA-RBP binding sites at nucleotide resolution.
• To improve upon existing methods by incorporating positional information and feature interactions.

Main Methods:

• Developed circdpb, a deep learning framework utilizing one-hot encoding and Gaussian-modulated position encoding for circRNA sequences.
• Employed dilated convolution feature pyramid (DCFP) and bidirectional gated recurrent unit (BiGRU) for enhanced feature extraction.
• Validated the model on 37 benchmark datasets through comparative experiments and ablation studies.

Main Results:

• circdpb demonstrates robust performance in predicting circRNA-RBP binding sites with nucleotide-level precision.
• The framework successfully identified known motifs, aligning with existing biological knowledge.
• Feature visualization confirmed the predictive capabilities of circdpb.

Conclusions:

• circdpb offers a significant advancement in predicting circRNA-RBP binding sites with high accuracy.
• The model's ability to capture sequence and positional features enhances understanding of RNA-protein interactions.
• circdpb provides a valuable tool for research into diseases involving circRNA-RBP dysregulation.