AttentionCCA: An attention-based canonical correlation analysis framework for integrative multi-omics cancer...

Area of Science:

• Computational biology
• Bioinformatics
• Cancer genomics

Background:

• Cancer is a complex disease with significant molecular diversity.
• Multi-omics profiling offers insights but integrating heterogeneous data is challenging.
• Traditional correlation methods struggle with nonlinear, cross-modal dependencies.

Purpose of the Study:

• To develop a novel framework for integrating multi-omics data in cancer research.
• To improve the interpretability and biological relevance of multi-omics data integration.
• To enhance cancer subtype stratification and identify molecular signatures.

Main Methods:

• Introduced AttentionCCA, a supervised framework extending Canonical Correlation Analysis (CCA).
• Employed feature and cross-modal attention modules for adaptive reweighting and latent space alignment.
• Validated on TCGA datasets (BRCA, LGG, STAD, UCEC) and external datasets (METABRIC, CGGA, GSE62254, GSE17025).

Main Results:

• AttentionCCA outperformed traditional classifiers and integration baselines, achieving high accuracy (up to 0.93) and Macro-F1 (up to 0.92).
• Ablation studies confirmed the necessity of both attention modules for performance.
• Attention heatmaps identified biologically relevant pathways, including RNA splicing (STAD) and MAPK signaling (UCEC).

Conclusions:

• AttentionCCA effectively integrates multi-omics data, transforming complexity into interpretable structure.
• The framework enables robust cancer subtype stratification and reveals biologically grounded molecular signatures.
• Attention-guided representation learning enhances the understanding of cancer heterogeneity.