Multi-View State-Space Modeling for Zero-Shot circRNA-miRNA Interaction Prediction

CircRNA-miRNA interactions (CMIs) play regulatory roles in human diseases and provide clues for therapeutic target discovery. However, most existing prediction methods rely on circRNA-miRNA pairs and overlook disease-level regulatory contexts, which limits their ability to characterize cross-pathway dependencies in zero-shot inference.

We propose MV-S2M, a multi-view state-space modeling framework that integrates molecular attributes to enhance relational reasoning, enabling zero-shot CMI prediction. Specifically, MV-S2M performs structure-aware diffusion to enrich molecular priors and constructs disease-mediated semantic subgraphs. It then incorporates state-space modeling to capture long-range dependencies and view-specific regulatory patterns. The learned multi-view representations are fused through a linear projection and fed into an MLP for CMI prediction. Extensive experiments on multiple benchmark datasets demonstrate that MV-S2M achieves state-of-the-art performance under both zero-shot and conventional settings. Cold-start evaluations reveal distinct generalization behavior across different unseen-entity settings. Case studies illustrate the contribution of meta-path-guided semantic subgraphs to CMI prediction.

The code and data are publicly available at https://github.com/591286260/MV and https://doi.org/10.5281/zenodo.22858403.