Clustering on hierarchical heterogeneous data with prior pairwise relationships

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Traditional clustering methods overlook feature differences, limiting applications in complex biological data.
• Unequal feature treatment in cancer data analysis hinders accurate diagnosis and effective anti-cancer therapies.
• Heterogeneity in biological data and cancer itself necessitates advanced clustering approaches.

Purpose of the Study:

• To propose a hierarchical clustering framework for heterogeneous data incorporating prior pairwise relationships.
• To characterize feature differences and identify hierarchical structures through rough and refined clustering.
• To enhance cancer subtyping and provide deeper biological insights beyond existing methods.

Main Methods:

• Developed a clustering framework for hierarchical heterogeneous data.
• Employed rough clustering for initial groupings and refined clustering for subtype identification.
• Integrated prior pairwise relationships of samples to improve clustering performance.

Main Results:

• The refined clustering successfully identified distinct cancer subtypes, offering deeper insights.
• The framework demonstrated flexibility in incorporating prior information, boosting clustering accuracy.
• Statistical consistency properties, including parameter estimation and structure determination, were rigorously established.

Conclusions:

• The proposed method outperforms existing approaches in simulation studies, especially with added prior information.
• Hierarchical clustering revealed necessary and reasonable structures in lung adenocarcinoma analysis using imaging and omics data.
• This approach offers a more nuanced understanding of cancer heterogeneity and potential therapeutic targets.