Multi-task adaptive deep sparse canonical correlation analysis for multi-omics cancer survival prediction.

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Integrating multi-omics data (DNA methylation, mRNA expression) is crucial for understanding cancer progression.
• Existing methods often model omics layers independently or use linear assumptions, missing nonlinear cross-omics interactions.

Purpose of the Study:

• To develop a novel framework, MT-ADSCCA, for joint multi-omics data integration and cancer survival prediction.
• To capture nonlinear cross-omics structures and identify interpretable multi-omics biomarkers.

Main Methods:

• Proposed MT-ADSCCA: a multitask adaptive deep sparse canonical correlation analysis framework.
• Employed nonlinear encoder architecture with sparse CCA and uncertainty-guided adaptive weighting.
• Utilized a BiLSTM-Cox survival network with chromosomally ordered genes for prognosis.

Main Results:

• MT-ADSCCA achieved superior concordance indices across BRCA, GBMLGG, and KIPAN TCGA cohorts.
• Outperformed six feature-selection and four survival-model baselines in prediction accuracy.
• Demonstrated clear risk group separation via Kaplan-Meier analysis and biologically enriched features.

Conclusions:

• MT-ADSCCA offers a robust and interpretable framework for multi-omics integration in cancer research.
• The framework effectively captures nonlinear cross-omics dependencies for improved survival prediction.
• Selected biomarkers are biologically relevant, enhancing the interpretability of cancer prognosis models.