Clusternomics: Integrative context-dependent clustering for heterogeneous datasets.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Integrative clustering analyzes multiple datasets (gene expression, methylation) to group biological samples.
• Existing methods often assume consistent cluster structures across all datasets, which is not always realistic.
• Heterogeneous biological data can exhibit varied structures, with clusters merging or separating across datasets.

Purpose of the Study:

• To develop a probabilistic clustering method, Clusternomics, for identifying sample groups across heterogeneous datasets with non-consistent cluster structures.
• To model local cluster assignments within individual datasets and extract global structure from these assignments.
• To apply Clusternomics for cancer subtyping and identify clinically meaningful patient groups.

Main Methods:

• Developed Clusternomics, a probabilistic clustering algorithm using a hierarchical Dirichlet mixture model.
• Modeled clusters at both local (individual dataset) and global (across datasets) levels.
• Evaluated performance on simulated data with varying common structure and real-world cancer datasets (TCGA).

Main Results:

• Clusternomics outperformed existing integrative and consensus clustering algorithms on simulated data.
• Applied to TCGA breast cancer data (gene expression, miRNA, methylation, proteomics), it identified clinically meaningful subtypes with distinct survival probabilities.
• Demonstrated scalability and clinically significant results on high-dimensional TCGA lung and kidney cancer datasets.

Conclusions:

• Clusternomics effectively identifies sample groups with shared global behavior across heterogeneous datasets, even with varying local cluster structures.
• The method provides a robust approach for integrative clustering, particularly valuable for complex biological data and cancer subtyping.
• Clusternomics offers a scalable and clinically relevant tool for analyzing multi-omics data to discover patient subgroups.