CircMAN: Multi-channel Attention Networks Based on Feature Fusion for CircRNA-Binding Protein Site Prediction

Circular RNAs interact with RNA-binding proteins, participate in and regulate gene transcription and expression, and play an indispensable role in the pathogenesis of various diseases, especially neurodegenerative diseases and cancers. Given that traditional biological experiments are often time-consuming and costly, developing computational methods for predicting circRNA-binding protein sites is crucial. Current computational methods for extracting circRNA sequence features are relatively simple, unable to fully express the information of circRNA, and also struggle to fully extract information from features. Therefore, we propose the CircMAN method, which employs six methods for extracting circRNA sequence features to aggregate global and local sequence information using the multi-scale channel attention module. Then, the internal dependencies of the six types of features are learned using bidirectional gated recurrent units and self-attention. The CircMAN method achieved an AUC of 0.923 on 37 circRNA datasets. Its effectiveness was confirmed through comparisons with five different methods and multiple ablation experiments. The source code can be available at https://github.com/lhlmeng/CircMAN .