IRESeek: structure-informed deep learning method for accurate identification of internal ribosome entry sites in circular RNAs

The internal ribosome entry site (IRES) is a special type of RNA <i>cis</i>-acting element that can initiate translation independently of the 5' cap structure and is widely found in viral RNAs and eukaryotic messenger RNAs. In recent years, an increasing number of studies have revealed that IRES elements also exist in circular RNAs (circRNAs) and mediate their translation. CircRNAs exhibit high stability and tissue specificity, playing critical roles in various physiological and pathological processes. Their coding potential provides important clues for the discovery of novel functional proteins. However, due to the nonlinear structure of circRNAs and the complexity of IRES-mediated regulatory mechanisms, accurately identifying IRES elements within circRNAs remains a significant challenge. Here, we propose IRESeek, a dual-branch deep learning framework for highly accurate detection of IRES elements in circRNA, which utilizes transformer for RNA sequence modeling and graph convolutional network for RNA structural guidance. To grasp the structural patterns of circRNAs, IRESeek employs physical-based thermodynamic energy of RNA secondary structure-base pair motif energy and the base pair probability as guidance structural characteristics to incorporate with RNA sequence, enabling comprehensive joint learning of RNA sequence and base pair interactions.