Abstract Despite significant progress in miRNA–disease association prediction, accurate inference remains hampered by the extreme sparsity of interaction networks, the heterogeneity of biological data modalities, and the redundancy introduced by simple feature concatenation. These limitations are especially pronounced in major cancers such as lung, breast, colorectal, gastric, and liver, where similarity-based methods often fail to capture critical pathological characteristics, underscoring the need to integrate histopathological images for richer disease representation. To address these challenges, we propose EMGMDA, which applies residual GraphSAGE to the sparse miRNA–disease graph to generate robust neighbor-aggregated embeddings, employs a nonlinear adaptive fusion module to learn high-order miRNA–disease feature interactions and eliminate redundancy, integrates multi-scale histopathological features from The Cancer Genome Atlas (TCGA) whole-slide images via a pretrained ResNet-18 and cross-attention mechanism to heighten sensitivity to tumor heterogeneity, and leverages triplet contrastive learning to refine the embedding space by drawing true associations closer and separating unrelated pairs, thereby improving discrimination in data-scarce scenarios. Experiments show that EMGMDA achieves an AUC of 0.9641 and an AUPRC of 0.9599 on HMDD v2.0, and further elevates performance to an AUC of 0.9742 and an AUPRC of 0.9719 on HMDD v3.2, significantly surpassing state-of-the-art methods. Case studies on lung, esophageal, breast, and colorectal cancers further validate its reliability and practical utility.
Sui et al. (Tue,) studied this question.