iCluster

iCluster is a statistical framework for integrative clustering of multiple genomic data types. It identifies subgroups of samples by inferring a shared latent structure that accounts for variation across data modalities such as gene expression, DNA methylation, copy number variation, and somatic mutation data.

Methodology: The original model is a joint latent variable model where each data block is modeled as a linear function of a small number of latent variables plus block-specific noise. Sparsity is encouraged via L1 penalties on the loading matrices to select features relevant to the cluster structure. The latent variables capture the common signal that drives clustering, while block-specific noise accounts for differences among data types. Estimation is typically performed by an alternating optimization scheme, often framed within a penalized likelihood or Bayesian EM approach, and the resulting sample clusters are derived from the latent representations.

Extensions: iClusterPlus generalizes to non-Gaussian data using exponential family distributions, enabling binary and count data; it also supports missing values, different distributional assumptions, and multiple penalty choices.

Applications: Used for cancer subtyping and integrative analyses across cohorts, helping to identify molecularly defined subtypes and to interpret relationships among genomic layers. Datasets commonly include expression, methylation, and copy number data.

Implementation: R packages such as iClusterPlus provide user-accessible implementations, along with tutorials and documentation. The method has become a reference for multi-omics integrative clustering in bioinformatics.

Limitations: Requires careful normalization and tuning of the number of clusters and sparsity parameters; results can be sensitive to model assumptions and data quality; computationally intensive on large datasets.