CAMDA: Contrastive and meta-path aware learning for miRNA-disease association prediction
Abstract
The identification of miRNA-disease associations (MDAs) is fundamental to elucidating complex disease pathophysiology. Computational methods for MDA prediction often face challenges in integrating diverse semantic relationships and learning robust representations from sparse data. To address these issues, we propose CAMDA, a framework that learns node representations from multiple meta-path-defined graph views. CAMDA utilizes a tri-perspective graph convolutional network architecture to process these views through three parallel pathways, generating specific, consensus, and topology-aware embeddings. To optimize these representations, the framework employs a dual-objective contrastive learning mechanism based on mutual information maximization. This mechanism operates at both the nodal level, to preserve local meta-path-specific information, and at the global level, to capture consistent patterns across the network. The final representations are generated by concatenating the embeddings from all three perspectives for downstream prediction tasks. Experiments on the HMDD v3.2 dataset demonstrate that CAMDA achieves an average AUC of 95.67% under five-fold cross-validation. Case studies on three gastrointestinal cancers (esophageal, gastric, and colorectal neoplasms) confirm the model's ability to identify biologically relevant associations, with validation in established databases supporting the predicted novel miRNA-disease associations.