Graph neural network-based discovery of miRNA-mRNA regulatory interactions in breast cancer

Breast cancer mortality is linked to dysregulation of miRNA-mediated gene regulation. This study developed a Heterogeneous Graph Transformer (HGT) link-prediction framework integrating miRTarBase interactions with TCGA-BRCA (and TCGA-OV to expand coverage). The model achieved accuracy 82%, AUC-ROC 90%, AUC-PRC 80%, F1 75%, and MCC 62%. High-confidence novel predictions included hsa-miR-4775 targeting SMAD4, ABCA1, and ZMAT3, and hsa-miR-520d-5p targeting BMPR2—targets implicated in TGF-β signalling, EMT, tumour suppression, and p53-mediated apoptosis. Graph-based deep learning can accelerate discovery of miRNA-mediated mechanisms for experimental validation.