circMAC: microRNA-conditioned binding-site localization on circular RNA isoforms.

Abstract

Circular RNAs (circRNAs) regulate gene expression in part through interactions with microRNAs (miRNAs), but identifying miRNA binding sites on full-length circRNA isoforms remains challenging. Binding context can differ across circRNA isoforms, and sequence continuity across the back-splice junction may be overlooked when circRNAs are represented as linear transcripts. Existing circRNA-miRNA resources and computational approaches mainly support association-level prediction or rule-based candidate-site screening. Although rule-based tools can be applied to full-length circRNA sequences, they do not directly learn miRNA-conditioned nucleotide-level binding-site localization while preserving circular sequence continuity. Here, we formulate circRNA-miRNA binding-site prediction as a miRNA-conditioned sequence-labeling task and present circMAC, a purpose-built framework for nucleotide-level localization on full-length circRNA isoforms. Given a full-length circRNA isoform and a mature miRNA sequence, circMAC predicts a binding probability for each circRNA nucleotide. circMAC combines established sequence-modeling components in a task-specific architecture, including attention-based global context modeling, Mamba-based sequential modeling, and convolutional local motif extraction. Paired miRNA information is incorporated through cross-attention, allowing each circRNA nucleotide to be evaluated in a miRNA-specific context. We evaluated circMAC against pretrained RNA language models, conventional sequence encoders, alternative pretraining strategies, architectural ablations, and stricter isoform-disjoint and back-splice-junction-disjoint splits. circMAC showed improved nucleotide-level localization performance under the evaluated benchmark settings, while qualitative and aggregate analyses indicated concentration of prediction signals around annotated binding-site regions. These results support task-specific full-length circular isoform modeling for prioritizing candidate circRNA-miRNA binding-site regions.