Router-guided dual attention with graph masked autoencoder for microRNA-drug association prediction

MicroRNAs (miRNAs), a major class of small non-coding RNAs, regulate gene expression and serve as key mediators of drug response. Identifying miRNA-drug associations is therefore important for elucidating drug-response mechanisms and supporting therapeutic discovery. However, existing computational methods still struggle to simultaneously capture long-range dependency patterns and fine-grained local topological structures, particularly under sparse association observations and noisy biological data. To address this challenge, we present router-guided dual attention with graph masked autoencoder (RGDAGMAE), a unified predictive framework that couples route-mediated global dependency modeling with gated graph masked reconstruction. The router-guided dual attention module couples softmax-based routed aggregation with low-rank route-mediated interaction to model global dependencies with controlled computational complexity. In parallel, the gated graph masked autoencoder performs self-supervised feature masking and gate-controlled message passing to reconstruct local structures and suppress spurious noise. Their complementary representations are then fused to generate association scores. Extensive experiments on three benchmark datasets show that RGDAGMAE consistently outperforms representative competing methods. Further ablation studies confirm the effectiveness of adaptive attention routing and masked graph reconstruction in capturing informative miRNA-drug association patterns. Case studies on vorinostat and miR-509-3p further support the biological plausibility of RGDAGMAE in prioritizing candidate miRNA-drug associations.