This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: Circular RNAs (circRNAs) are loop-shaped molecules that can influence gene regulation through interactions with microRNAs (miRNAs). This matters because circRNAs can exist as different full-length isoforms, and their back-splice junction (BSJ) connects the two ends of the RNA.
Treating a circular RNA like a linear transcript can therefore overlook sequence information near this junction. Giltae Song, a professor in the School of Computer Science and Engineering at Pusan National University, South Korea, and his colleagues developed circMAC, an AI bioinformatics tool for predicting candidate microRNA binding sites on circular RNAs. The framework takes a full-length circRNA isoform and a mature miRNA as inputs and predicts a binding probability for every circRNA nucleotide.
This paper was published in Briefings in Bioinformatics on Sept. 9, 2026. "The key advance is moving beyond asking whether a circRNA and miRNA interact to asking where that interaction is likely to occur," said Song. To preserve the distinctive circular sequence, circMAC combines three complementary sequence-modeling approaches.
Attention captures information from across the circRNA, Mamba models sequential dependencies, and a one-dimensional convolutional neural network identifies short local sequence patterns associated with binding sites. Circular padding in the convolutional branch allows patterns spanning the BSJ to be considered continuously rather than interrupted at the sequence boundary. The model also uses cross-attention to incorporate information from the specific miRNA being analyzed, allowing each circRNA nucleotide to be evaluated in its miRNA-specific context.
It also uses structure-aware pretraining based on base-pairing prediction. The researchers evaluated circMAC using full-length circRNA isoforms supported by AGO-CLIP binding-site annotations. The main pair-disjoint data set contained 19,674 training pairs and 7,838 test pairs, with 5.7% of nucleotide positions labeled as binding sites.
At the 1,000-nucleotide setting, circMAC with base-pairing prediction pretraining achieved an F1-macro score of 0.762, an AUROC of 0.902 and an AUPRC of 0.545, outperforming the evaluated pretrained RNA language models. It also showed the strongest overall performance among evaluated full-length models under isoform-disjoint and BSJ-disjoint tests. "By retaining full-length sequence context and conditioning predictions on the specific miRNA, the framework can provide more focused candidate regions for follow-up experiments," said Song. circMAC can help computational biology researchers identify potential microRNA binding regions on circular RNA at nucleotide resolution, enabling more targeted experimental validation.
Extract — continue reading at the source.
