Clustering Sparse Data With Feature Correlation With Application to Discover Subtypes in Cancer

Area of Science:

• Computational biology
• Bioinformatics
• Network science

Background:

• High-dimensional data presents challenges for traditional similarity calculations.
• Existing methods often rely on sample networks, neglecting feature relationships.
• Data sparseness is a significant hurdle in analyzing complex biological datasets.

Purpose of the Study:

• To develop a novel network-based similarity metric for high-dimensional data.
• To incorporate feature interaction networks to overcome data sparseness.
• To improve the accuracy and scope of cancer subtype discovery.

Main Methods:

• Proposed a novel network-based similarity metric incorporating feature interaction networks.
• Introduced a Feature Alignment Similarity measure projecting samples into a feature network.
• Applied the metric to tumor mutational data using gene interaction networks.

Main Results:

• The proposed metric effectively measures similarity even without shared features, based on network proximity.
• Demonstrated superior performance in cancer subtype discovery compared to top competitors.
• Identified novel cancer subtypes missed by traditional clustering algorithms.

Conclusions:

• The novel similarity metric successfully leverages feature interaction networks to improve sample similarity calculations.
• This approach offers a powerful tool for uncovering complex patterns in high-dimensional biological data.
• The method significantly advances cancer subtype discovery, revealing previously undetected subtypes.