Prediction of Circular RNA-RBP Binding Sites Based on Multi-source Features and Cascade Forest

Circular RNAs (circRNAs) are a non-coding RNAs with a special circular structure. They play an important role in gene regulation and interact with RNA-binding proteins (RBPs) to generate binding sites. There is growing evidence that it is essential to predict the interactions between circRNAs and RBP binding sites for diagnosing diseases and providing a potential target to treat diseases. In this paper, we propose the DFCRBP to identify the binding sites of circRNA-RBPs. The DFCRBP is composed of feature encoding module, deep features extraction module and classifier. In feature encoding module, we adopt five feature encoding methods for circRNA sequences. The deep feature extraction module consists of two parts: Joint representation learning module (JRLM) and MSA Transformer. We input five encoded features into the JRLM module to extract local features. CircRNA sequences are fed into MSA Transformer to extract global features. The obtained features are fed into deep forest with cascade forest structure to predict circRNA-RBP binding sites. To validate the effectiveness of the DFCRBP, we compared its performance on 12 circRNA-RBP datasets with existing methods. The experimental results demonstrate that our method achieves excellent binding sites prediction.