Assessment of data transformations for model-based clustering of RNA-Seq data

Quality control, global biases, normalization, and analysis methods for RNA-Seq data are quite different than those for microarray-based studies. The assumption of normality is reasonable for microarray based gene expression data; however, RNA-Seq data tend to follow an over-dispersed Poisson or negative binomial distribution. Little research has been done to assess how data transformations impact Gaussian model-based clustering with respect to clustering performance and accuracy in estimating the correct number of clusters in RNA-Seq data.

Here, researchers from the University of Kansas Medical Center investigate Gaussian model-based clustering performance and accuracy in estimating the correct number of clusters by applying four data transformations (i.e., naïve, logarithmic, Blom, and variance stabilizing transformation) to simulated RNA-Seq data. To do so, an extensive simulation study was carried out in which the scenarios varied in terms of: how genes were selected to be included in the clustering analyses, size of the clusters, and number of clusters. Following the application of the different transformations to the simulated data, Gaussian model-based clustering was carried out. To assess clustering performance for each of the data transformations, the adjusted rand index, clustering error rate, and concordance index were utilized. As expected, our results showed that clustering performance was gained in scenarios where data transformations were applied to make the data appear “more” Gaussian in distribution.

Simulation schematic

Negative binomial (NB) parameters were obtained from 100 top genes and 100 randomly selected genes based upon Median Absolute Deviation (MAD) of expression values taken from ovarian cancer RNA-Seq tumors (N = 55 patients). Data were then simulated to reflect varying cluster sizes, equal and unequal, for K = 1 (i.e., no clusters), 2, and 3 clusters using the NB parameters. One hundred datasets were simulated for four parent dataset categories which reflected gene selection and cluster size (1) Top 100 genes with equal cluster sizes (TE); 2) Top 100 genes with unequal cluster sizes (TX); 3) Random 100 genes with equal cluster Sizes (RE); and 4) Random 100 genes with unequal cluster sizes (RX)). Data transformations and model-based clustering were applied to all datasets and evaluated according to normality measures and clustering evaluation metrics.