Prediction of Breast Cancer Metastasis Using Graph Neural Networks: A Comprehensive Study

Abstract
Early prediction of breast cancer metastasis is crucial for timely intervention and improved patient outcomes. In this study, we investigate Graph Neural Network (GNN) models for predicting metastasis in breast cancer by integrating high-dimensional molecular and clinical data. We leverage two large-scale breast cancer cohorts – METABRIC (∼2,000 tumors) and TCGA-BRCA (1,082 tumors) – which provide multi-omics profiles (gene expression, copy number, etc.) and metastasis status. After preprocessing (normalization, feature selection, and construction of patient similarity graphs), we train several GNN variants (Graph Convolutional Network, Graph Attention Network, and GraphSAGE) to classify patients as metastatic vs. non-metastatic. Models are evaluated using 5-fold cross-validation with metrics including accuracy, AUC, sensitivity, and specificity. Our best GNN model achieves an AUC of ~0.95 and accuracy ~93%, outperforming baseline methods (random forest, SVM, logistic regression) which yield AUCs around 0.85–0.90. Visualization of results via ROC curves confirms superior discriminative power for GNNs. We discuss how the graph-based approach captures complex inter-patient relationships and interactions between genes, leading to improved prediction. We also analyze model interpretability and limitations (e.g., need for graph construction, hyperparameter tuning). This comprehensive study demonstrates that GNNs offer significant advantages for metastasis risk prediction in breast cancer, laying the groundwork for future multimodal graph-based prognostic models.