Domain adaptation, self-supervision, and generative augmentation enhance GNNs for breast cancer prediction

Breast cancer presents substantial molecular heterogeneity, requiring accurate subtype classification, receptor-status prediction, and survival estimation for precision care. Existing machine-learning models often fail to generalize across cohorts or adapt to rare subtypes. We propose a unified graph neural network (GNN) framework that integrates multi-task learning, domain-adversarial adaptation, contrastive self-supervision, few-shot meta-learning, and generative augmentation. Gene-expression data from TCGA-BRCA (1084 samples) and METABRIC (1980 samples) were mapped onto gene-centric PPI graphs and patient-similarity graphs. A shared encoder (including Graph Transformer variants) jointly predicts intrinsic subtypes (Luminal A, Luminal B, HER2-enriched, Basal-like), ER/PR/HER2 biomarkers, and overall survival (OS) using a Cox proportional hazards head. Validation included fivefold cross-validation and strict TCGA → METABRIC transfer testing. The multi-task Graph Transformer achieved subtype F1 = 0.872, ER/PR/HER2 AUCs of 0.960/0.943/0.918, and C-index = 0.721. Domain adaptation improved external subtype F1 from 0.738 to 0.801. For the HER2-enriched subtype, MAML enabled few-shot prediction with F1 = 0.782, while MolGAN augmentation increased HER2 AUC to 0.935. GNNExplainer highlighted biologically consistent drivers, including ESR1, ERBB2, and PGR, aligning with known hormonal and HER2 signaling mechanisms. This study introduces a comprehensive, interpretable GNN framework that unifies subtyping, biomarker prediction, and survival modeling while improving cross-cohort robustness and rare-subtype adaptation. The combination of multi-task learning, domain adaptation, self-supervision, and generative augmentation demonstrates strong potential for clinically actionable decision support.