Although a large number of clustering algorithms have been proposed to identify groups of co-expressed genes from microarray data, the question of if and how such methods may be applied to RNA-seq data remains unaddressed. In this work, researchers from Université Paris-Saclay and Université de Toulouse investigate the use of data transformations in conjunction with Gaussian mixture models for RNA-seq co-expression analyses, as well as a penalized model selection criterion to select both an appropriate transformation and number of clusters present in the data. This approach has the advantage of accounting for per-cluster correlation structures among samples, which can be quite strong in real RNA-seq data. In addition, it provides a rigorous statistical framework for parameter estimation, an objective assessment of data transformations and number of clusters, and the possibility of performing diagnostic checks on the quality and homogeneity of the identified clusters. The researchers analyze four varied RNA-seq datasets to illustrate the use of transformations and model selection in conjunction with Gaussian mixture models. Finally, they propose an R package coseq (co-expression of RNA-seq data) to facilitate implementation and visualization of the recommended RNA-seq co-expression analyses.
Evaluation of clustering quality for the Graveley et al. (top) and Fietz et al. mouse (bottom) data. (left) Maximum conditional probabilities τmax(i) for each cluster, sorted in decreasing order by the cluster median. (right) Barplots of cluster sizes, according to τmax(i) greater than or less than 0.8, sorted according to the number of genes with τmax(i) > 0.8.
Availability – https://github.com/andreamrau/coseq