Low-Rank Joint Subspace Construction for Cancer Subtype Discovery

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Multimodal data integration is crucial for cancer subtype discovery, but selecting relevant data types and handling high-dimensional, low-sample data are significant challenges.
• Existing methods often struggle with the 'high dimension-low sample size' issue inherent in individual data modalities.

Purpose of the Study:

• To propose a novel algorithm for constructing a low-rank joint subspace from individual modalities for improved cancer subtype discovery.
• To address the challenges of modality selection and dimensionality reduction in multimodal cancer data analysis.

Main Methods:

• Developed a novel algorithm to create a low-rank joint subspace from individual low-rank subspaces of high-dimensional modalities.
• Employed statistical hypothesis testing for accurate rank estimation and signal-noise separation.
• Introduced two quantitative indices to assess modality relevance and shared cluster information.
• Utilized subspace intersection for noise filtering and cluster information selection during data integration.

Main Results:

• The proposed algorithm effectively selects relevant modalities with maximum shared information for joint subspace construction.
• Clustering on the extracted joint subspace yielded subtypes with higher resemblance to clinically established subtypes compared to existing methods.
• Survival analysis confirmed significant differences in survival profiles among the identified subtypes.
• Robustness analysis demonstrated the stability of the identified subtypes against data perturbations.

Conclusions:

• The novel algorithm provides an effective framework for multimodal data integration in cancer subtype discovery.
• The identified subtypes exhibit strong clinical relevance and prognostic value.
• The method demonstrates superior performance and robustness over existing integrative clustering approaches.