Heterogeneous dual-channel and interpretable graph representation learning with global virtual nodes for microRNA-mediated drug sensitivity prediction.

Drug sensitivity critically affects therapeutic outcomes, and microRNAs (miRNAs) play a key role in regulating drug response by modulating genes involved in drug metabolism and action. However, existing computational methods for predicting miRNA-drug sensitivity associations are often limited by heterogeneous network structures and severe data sparsity, which hinder effective feature propagation and robust learning. To address these challenges, we propose HDIGRL, a channel-aware heterogeneous graph representation learning framework centered on channel-gated global heterogeneous propagation for miRNA-mediated drug sensitivity prediction. HDIGRL models miRNAs and drugs from complementary structural and interaction-derived perspectives via a dual-channel feature extraction strategy. HDIGRL introduces channel-gated global heterogeneous propagation, in which global virtual nodes first enable graph-level context exchange and a channel-wise propagation gate then recalibrates propagated embeddings to emphasize discriminative feature channels and suppress noisy or redundant ones. In addition, an imbalance-aware focal loss is adopted to improve robustness under extreme class imbalance. Experimental results on public datasets demonstrate that HDIGRL consistently outperforms existing methods, and further analyses reveal latent miRNA-mediated drug-sensitivity pathways, highlighting its potential for predictive modeling and biological interpretation.