Multi-Omics Graph Attention Network for Key Gene Prediction in Triple-Negative Breast Cancer.

Triple-negative breast cancer (TNBC) is an aggressive malignancy lacking effective targeted therapies, underscoring the need for robust and interpretable biomarkers for personalized treatment. Here, we propose a graph attention network (GAT)-based multimodal framework that integrates scRNA-seq, scATAC-seq, and radiomics to capture cross-modal regulatory interactions underlying TNBC heterogeneity. Transcriptional, chromatin accessibility, and imaging features are aligned via canonical correlation analysis, with intercellular communication-derived gene relationships and transcription factor-binding-guided edges incorporated into a multimodal graph. Multi-head attention enables adaptive weighting of omics-specific interactions, while an ensemble multilayer perceptron with variational dropout stratifies patient prognosis. The model demonstrates strong predictive performance in an external TCGA-TNBC cohort (log-rank p<0.01), outperforming single-omics and alternative graph-based approaches (AUC-ROC = 0.839, 95% CI: 0.81-0.87). Pathway analysis validates canonical TNBC drivers, including PI3K, ERBB2, PTK6, and EGFR signaling, while revealing previously underappreciated regulatory programs involving complement-coagulation cascades, ECM-integrin-focal adhesion signaling, leukocyte transendothelial migration, sphingolipid-mediated metabolic-immune coupling, and nanoparticle-receptor interactions. Collectively, this framework provides an interpretable strategy for multimodal biomarker discovery in TNBC, uncovering both established and novel therapeutic vulnerabilities and offering a scalable approach toward precision oncology.