• Prediction of Breast Cancer Metastasis Using Graph Neural Networks: A Comprehensive Study
  • Amir Hossein Rahmani,1,* Nadia Sepehrimoghaddam,2 Paria Qanbari,3
    1. Department of Biomedical Engineering, Rouzbahan Higher Education Institute, Sari, Iran
    2. Faculty of Dentistry, Islamic Azad University, Medical Sciences Branch, Tehran, Iran
    3. Department of Biomedical Engineering, Payame Noor University, Qazvin, Iran


  • Introduction: Breast cancer metastasis is a major cause of mortality among breast cancer patients, emphasizing the critical need for accurate early prediction to improve clinical outcomes. This study investigates the use of Graph Neural Networks (GNNs) to predict metastasis risk by modeling complex inter-patient relationships through multi-omics and clinical data integration.
  • Methods: We utilized two major breast cancer datasets, METABRIC and TCGA-BRCA, encompassing gene expression, copy number variations, and clinical features. Patient similarity graphs were constructed based on molecular profiles, and GNN models (Graph Convolutional Network, Graph Attention Network, and GraphSAGE) were trained to classify metastatic versus non-metastatic patients. Five-fold cross-validation was employed, and model performance was evaluated using accuracy, AUC, sensitivity, and specificity metrics.
  • Results: The GNN models, particularly GCN, achieved superior predictive performance compared to traditional machine learning models. The best GNN model achieved an AUC of approximately 0.95 and an accuracy of 93%, outperforming baseline methods such as random forest, SVM, and logistic regression, which showed AUCs between 0.85 and 0.90.
  • Conclusion: This study demonstrates that GNN-based modeling of relational patient data significantly enhances breast cancer metastasis prediction. The findings contribute to the advancement of predictive oncology by offering a more accurate and biologically informed approach. Future research could explore integrating additional omics layers and developing explainable GNN architectures to further improve clinical applicability.
  • Keywords: Breast cancer, metastasis prediction, graph neural networks (GNN), multi-omics, machine learning