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

Circular RNAs (circRNAs) influence gene activity and cell behavior by binding to specific RNA binding proteins (RBPs), which is crucial in various diseases. Few approaches achieve nucleotide-resolution accuracy in circRNA-RBP binding sites prediction, and they commonly overlook the positional information of nucleotides as well as the interaction between deep and shallow features within the model. Here, we propose circdpb, a deep learning-based framework, to predict circRNA-RBP binding sites with nucleotide-level precision. Firstly one-hot and Gaussian-modulated position encoding are integrated to represent circRNA sequences. circdpb then utilizes dilated convolution feature pyramid (DCFP) and bidirectional gated recurrent unit (BiGRU) to enhance feature extraction. On 37 benchmark datasets, comparative experiments and ablation studies suggest the robustness of circdpb. circdpb can also identify motifs which are aligned with existing known motifs. In addition, circdpb's predictive capability has been validated through visualization of features and binding sites.