caBIG VISDA: modeling, visualization, and discovery for cluster analysis of genomic data

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Existing genomic data clustering methods suffer from heuristic initialization, local optima, lack of cluster number detection, and inability to incorporate prior knowledge.
• The curse of dimensionality and confounding variables further complicate the discernment of meaningful cluster structures in genomic datasets.

Purpose of the Study:

• To develop the VIsual Statistical Data Analyzer (VISDA) for robust cluster modeling, visualization, and discovery in complex, high-dimensional genomic data.
• To address limitations of existing methods by incorporating hierarchical modeling, informative gene selection, and user guidance.

Main Methods:

• VISDA employs progressive, coarse-to-fine hierarchical clustering and visualization with hierarchical mixture modeling.
• It integrates supervised/unsupervised informative gene and data visualization, utilizing multiple projection methods for enhanced cluster structure revelation.
• Initialization and model order selection leverage Bayesian criteria and user-guided low-dimensional visualization subspaces for high-dimensional data.

Main Results:

• VISDA demonstrated robust and superior clustering accuracy compared to benchmark methods.
• The model order selection scheme proved effective for high-dimensional genomic data clustering.
• Biologically relevant gene clusters and pathological relationships among phenotypes were identified in muscular dystrophy and cancer datasets.

Conclusions:

• VISDA offers a powerful and flexible tool for gene, sample, and phenotype clustering in genomic analysis.
• The identified clusters provide valuable biological insights into diseases like muscular dystrophy and cancer at a molecular level.