Data Sheet 1_GCOA-Net: a graph-regularized cross-omics attention network for interpretable breast cancer molecular subtype classification.pdf

IntroductionAccurate intrinsic molecular subtyping is essential for precision management of breast cancer, yet multi-omics integration remains challenging due to high dimensionality, structured cross-omics dependencies, and the need for clinically interpretable and reliable predictions.
MethodsWe propose GCOA-Net, a graph-regularized cross-omics attention network that integrates transcriptomics, promoter-proximal DNA methylation, and miRNA expression. A biologically grounded heterogeneous graph connects CpG clusters to promoter-associated genes and miRNAs to their target genes. A relation-aware GNN encoder performs cross-omics message passing, while omics-specific and modality-level attention modules provide multi-level interpretability. We trained and evaluated models on TCGA-BRCA with repeated stratified five-fold cross-validation, benchmarking against classical early-fusion classifiers, integration frameworks, and deep multi-omics baselines. We additionally assessed ablations, subtype-specific explanations, robustness to missing modalities, calibration, and selective prediction.
ResultsGCOA-Net achieved the best overall performance (Acc = 0.912, Macro-F1 = 0.852, AUROC = 0.965) and improved calibration (ECE = 0.031) compared with baselines. Ablation analyses showed that biologically grounded cross-omics connectivity and graph regularization were key contributors, with degree-preserving edge randomization producing the largest performance drop. Attribution analyses identified subtype-consistent cross-omics biomarkers and compact explanatory subnetworks (e.g., ERBB2-centered regulation for HER2-enriched tumors). Under missing-modality scenarios, GCOA-Net degraded more gracefully and maintained better confidence reliability; selective prediction yielded a more favorable coverage–risk trade-off.
ConclusionHeterogeneous cross-omics graph modeling with graph regularization enables more accurate, robust, and interpretable breast cancer subtype classification, and provides a confidence-aware framework for molecular stratification that warrants further validation in independent multi-omics cohorts.