TransRM: Weakly supervised learning of translation-enhancing N6-methyladenosine (mA) in circular RNAs.

As our understanding of Circular RNAs (circRNAs) continues to expand, accumulating evidence has demonstrated that circRNAs can interact with microRNAs and RNA-binding proteins to modulate gene expression. More importantly, a subset of circRNAs has been reported to possess coding potential, enabling them to translate into functional proteins. Recent studies also indicate that the N6-methyladenosine (mA)-modified start codon may function as an Internal Ribosome Entry Site (IRES), influencing the translation of circRNAs. Therefore, elucidating how mA regulates circRNA translation potential could significantly advance circRNA research, including the development of circRNA-based vaccines. However, to our knowledge, there are currently no computational tools specifically designed for this purpose. To bridge this gap, we have developed the first computational model, termed TransRM, to predict the impact of base-resolution mA sites on circRNA translation. Our model employs weakly supervised learning with two convolution layers. These layers extract RNA modification features, and a bidirectional gated recurrent unit predicts the contribution of each RNA modification to circRNA translation. Subsequently, the RNA modification features are then integrated with their contribution to assess the probability of circRNA translation using a random forest algorithm. TransRM has demonstrated efficiency in identifying translation-enhancing mA sites, with an AUROC of 0.9188 and an AUPRC of 0.9371, respectively. We hope that our newly proposed model could help to broaden our understanding of circRNA regulation at the epitranscriptome layer, particularly in identifying translated circRNAs, thereby contributing to the development of more effective circular RNA-based therapeutics.