Clustering distributions with the marginalized nested Dirichlet process.

Area of Science:

• Computational biology
• Statistical modeling
• Bioinformatics

Background:

• Clustering gene expression data is crucial for understanding biological pathways.
• Existing methods for clustering distributions or gene-gene interaction patterns have limitations.
• Accurate inference of gene relationships is essential for disease research.

Purpose of the Study:

• To introduce a novel marginal nested Dirichlet process for clustering distributions and histograms.
• To apply this model for clustering genes based on gene-gene interaction patterns.
• To enable simulation-exact inference for improved clustering accuracy.

Main Methods:

• Developed a marginal version of the nested Dirichlet process.
• Applied the model to cluster genes using gene-gene interaction coefficients from the Zodiac database (TCGA data).
• Adjusted for covariates including copy number variation, methylation, and protein activation in an auto-logistic model.

Main Results:

• The proposed model allows for simulation-exact inference, unlike truncated approximations.
• Demonstrated favorable performance compared to k-means, truncated NDP, and hierarchical clustering.
• Successfully clustered genes related to DNA mismatch repair (DMR) based on interaction distributions.

Conclusions:

• The marginal nested Dirichlet process provides an effective and accurate method for clustering distributions, particularly for gene-gene interaction data.
• This approach offers a significant improvement over existing clustering techniques for biological data analysis.
• The model's ability to perform exact inference enhances its reliability in complex biological studies.