DAMFCMI: capturing cross-view interactions via hybrid attention for circRNA-MiRNA interaction prediction.

Motivation: Circular RNAs (circRNAs) and microRNAs (miRNAs) play pivotal roles in gene expression regulation, where understanding their interactions (CMIs) is essential for deciphering the molecular mechanisms behind cellular physiological and pathological states. Most existing approaches to CMI prediction are constrained by their reliance on shallow, single-view representations, while deep models typically align only on final embeddings, thereby neglecting the rich layer-wise interactions that are critical for capturing biological complexity. Results: To address these issues, we propose DAMFCMI, a novel method for CMI prediction. DAMFCMI characterizes circRNAs and miRNAs through three distinct feature views: sequence-based, attribute-based, and behavior-based features. A hybrid attention mechanism captures dependencies within individual views through multi-head self-attention and across views through cross-attention, enabling comprehensive modeling of feature interactions. Experimental results show that DAMFCMI outperforms state-of-the-art methods across three benchmark datasets. Visualization analyses demonstrate that the hybrid attention architecture enhances feature discriminability through effective multi-view feature integration. Moreover, case studies show that 13 out of 15 predicted CMIs predicted by DAMFCMI are supported by evidence in the PubMed literature, underscoring its potential for uncovering biologically relevant interactions. Availability and implementation: The data and source code are available at https://github.com/yadxbiolab/DAMFCMI.