CALT-GNN: a graph neural network with cross-attention and long-tail experts for multi-omics cancer subtype classification

Cancer subtype classification based on molecular features is important for characterizing tumor heterogeneity, evaluating patient prognosis, and supporting precision treatment. Multi-omics integration provides complementary molecular information across genomic, epigenomic, transcriptomic, and proteomic levels. However, the high dimensionality of multi-omics data, complex cross-omics relationships, and imbalanced subtype distributions remain major challenges. To address these issues, we propose CALT-GNN, a graph neural network framework that integrates patient-similarity graph learning, cross-attention-based omics integration, long-tail expert classification, and learnable prediction fusion. The framework first constructs omics-specific patient-similarity graphs, learns latent patient representations using graph convolutional networks, and obtains a unified graph through similarity network fusion. A cross-attention branch models complementary relationships between CNA and DNA methylation as source modalities and mRNA expression as the target modality. In parallel, a long-tail expert branch uses Major and Minor Experts with prototype-guided routing to perform class-distribution-aware classification. The outputs of the two branches are subsequently combined using a learnable global weight. Experiments on eight TCGA multi-omics cohorts showed that CALT-GNN achieved competitive or comparable classification performance relative to representative baseline methods and stable performance on Macro-F1 and MCC, which are sensitive to class imbalance. Ablation and subtype-level analyses further supported the complementary contributions of cross-omics interaction modeling and adaptive long-tail expert routing, although the magnitude of the improvements varied across cohorts and subtypes. These results suggest that CALT-GNN provides a methodological framework for multi-omics cancer subtype classification under complex cross-omics relationships and imbalanced subtype distributions.