DeepCIP: a multimodal deep learning method for the prediction of internal ribosome entry sites of circRNAs

MotivationCircular RNAs (circRNAs) have been found to have the potential to code proteins. Internal ribosome entry sites (IRESs) are key RNA regulatory elements for the translation of proteins by circRNAs through a cap-independent mechanism. IRES can be identified by bicistronic assay, but the method is time-consuming and laborious. Therefore, it is important to develop computational methods for facilitating IRES identification, evaluation, and design in circRNAs. ResultsIn this study, we proposed DeepCIP, a multimodal deep learning approach for circRNA IRES prediction, by exploiting both sequence and structure information. As far as we know, DeepCIP is the first predictor for circRNA IRESs, which consists of an RNA processing module, an S-LSTM module, a GCN module, a feature fusion module, and an ensemble module. The comparative studies show that DeepCIP outperforms other comparative methods and justify the effectiveness of the sequence model and structure model of DeepCIP for extracting features. We found that the integration of structural information on the basis of sequence information effectively improves predictive performance. For the real circRNA IRES prediction, DeepCIP also outperforms other methods. DeepCIP may facilitate the study of the coding potential of circRNAs as well as the design of circRNA drugs. DeepCIP as a standalone program is freely available at https://github.org/zjupgx/DeepCIP.