Description
Breast cancer is a highly heterogeneous disease that requires precise molecular subtype classification to support personalized treatment and prognosis. This research introduces an Attention-Based Graph Convolutional Network (AGCN), a supervised deep learning framework designed to integrate multi-omics data for robust breast cancer subtype identification.
The proposed AGCN model combines messenger RNA expression, DNA methylation, and copy number variation data with biological prior knowledge derived from protein–protein interaction networks. Multiple attention mechanisms are incorporated to capture cross-omics relationships and to adaptively weight the contribution of each molecular data type. Graph convolution layers further enhance feature learning by modeling gene–gene interactions within the biological network structure.
The model is evaluated on large-scale TCGA breast cancer datasets and demonstrates consistent improvements over traditional machine learning models and standard deep learning approaches. Performance gains are observed across key metrics, including AUC, accuracy, and Matthews correlation coefficient. The study further applies layer-wise relevance propagation to interpret model decisions, enabling the identification of subtype-specific and patient-specific gene markers with known biological relevance.
This paper is particularly valuable for researchers and practitioners in bioinformatics, cancer genomics, and medical artificial intelligence. It provides a well-validated and interpretable framework for multi-omics integration and molecular subtype classification in oncology.