DVCH-LMI: A dual-view contrastive framework for lncRNA-miRNA interaction prediction under class imbalance

Identifying long non-coding RNA (lncRNA) and microRNA (miRNA) interactions (LMIs) is essential for elucidating post-transcriptional regulation and disease-related molecular mechanisms. Accurate identification of LMIs can facilitate biomarker discovery, RNA-targeted therapeutics, and data-driven drug design. However, robust prediction and candidate prioritization remain challenging due to the limited number of experimentally validated interactions and the large number of unknown pairs. Accordingly, we propose DVCH-LMI, a dual-view contrastive heterogeneous representation learning framework. DVCH-LMI initializes node features with Doc2Vec embeddings trained on k-mer vectors, constructs an adjacency view and a similarity view over shared bipartite interaction semantics, and learns complementary embeddings using two heterogeneous encoders with identical architectures but independent parameters. A symmetric information noise-contrastive estimation (InfoNCE) loss is further introduced to achieve node-level cross-view alignment. The aligned embeddings are then fused, and interaction prediction is performed by a hybrid scorer composed of a similarity branch and a multi-layer perceptron (MLP) branch. To better align with candidate prioritization tasks in computational biology and therapeutic design, we integrate random negative sampling and hard negative sampling to enhance the model's candidate prioritization ability in an imbalanced candidate space. On an animal benchmark dataset, DVCH-LMI achieves AUC of 0.9465 and AUPR of 0.9468 under the 1:1 setting, and it maintains strong positive-oriented ranking performance under more imbalanced settings such as 1:20 and 1:50. Ablation studies verify the effectiveness of key components. Case studies on MALAT1 and NEAT1 further show that DVCH-LMI prioritizes literature-supported interactions. Cross-view alignment visualizations and structure-consistency analyses provide additional representational evidence for the interpretability and reliability of our model. Overall, DVCH-LMI serves as both an effective LMI predictive model and a robust candidate prioritization framework for data-sparse and highly class-imbalanced scenarios, with promising potential for RNA regulatory mechanism studies and RNA-guided drug design.