The discovery of transcriptional modules by a two-stage matrix decomposition approach

We address the problem of identifying gene transcriptional modules from gene expression data by proposing a new approach. Genes mostly interact with each other to form transcriptional modules for context-specific cellular activities or functions. Unraveling such transcriptional modules is important for understanding biological network, deciphering regulatory mechanisms and identifying biomarkers.

The proposed algorithm is based on two-stage matrix decomposition. We first model microarray data as non-linear mixtures and adopt the non-linear independent component analysis to reduce the non-linear distortion and separate the data into independent latent components. We then apply the probabilistic sparse matrix decomposition approach to model the 'hidden' expression profiles of genes across the independent latent components as linear weighted combinations of a small number of transcriptional regulator profiles. Finally, we propose a general scheme for identifying gene modules from the outcomes of the matrix decomposition.

The proposed algorithm partitions genes into non-mutually exclusive transcriptional modules, independent from expression profile similarity measurement. The modules contain genes with not only similar but different expression patterns, and show the highest enrichment of biological functions in comparison with those by other methods. The usefulness of the algorithm was validated by a yeast microarray data analysis.

The software is available upon request to the authors.