Related Experiment Video
Updated: Aug 5, 2026
Multiomics Analysis of TMEM200A as a Pan-Cancer Biomarker
Published on: September 15, 2023
AttentionCCA: An attention-based canonical correlation analysis framework for integrative multi-omics cancer
Guangji Zhang1, Chunxiao Zhang1, Heng Zhang1
1Department of Biomedical Engineering, School of Control Science and Engineering, Shandong University, Jinan, Shandong, 250061, China.
Background:
Cancer is not a single disease but a dynamic ecosystem of molecular diversity. Multi-omics profiling provides a panoramic view of tumor biology, yet integrating these heterogeneous modalities into coherent, interpretable representations remains a fundamental challenge. Classical correlation-based methods capture shared variation but fail to model nonlinear, cross-modal dependencies essential for biological discovery.
Methods:
We introduce AttentionCCA, a supervised framework that extends canonical correlation analysis (CCA) with feature- and cross-modal attention to jointly learn discriminative and biologically meaningful projections. AttentionCCA incorporates a weighted multi-class sparse CCA backbone within dual attention modules: feature attention adaptively reweights intra-modality signals, while cross-modal attention enables inter-omics dialogue by aligning latent spaces. The model was tested across four TCGA cancer types (BRCA, LGG, STAD, UCEC) integrating mRNA, DNA methylation, and miRNA data, and externally validated on METABRIC, CGGA, GSE62254, and GSE17025.
Results:
Across all datasets, AttentionCCA consistently outperformed traditional classifiers (KNN, SVM, Random Forest) and state-of-the-art integration baselines (DIABLO, MOGONET, bPLSDA), achieving accuracies up to 0.93 and Macro-F1 up to 0.92. Ablation studies showed that both attention modules are essential for performance gains, while attention heatmaps identified biologically relevant pathways such as RNA splicing in STAD and MAPK signaling in UCEC.
Conclusions:
By combining classical correlation analysis and modern attention mechanisms, AttentionCCA transforms multi-omics complexity into interpretable structure, enabling robust cancer subtype stratification and revealing biologically grounded molecular signatures. This work demonstrates how attention-guided representation learning can enhance our understanding of cancer heterogeneity.
