Hi,

 I am facing some difficulties with data transformations of my single cell RNA-Seq data analyzed using DESeq2. This data set is different from typical RNA-Seq experiments.. For example, there is a subset of genes which will be present in one group and totally absent in the other, unlike typical data sets where down regulated genes will still be expressed at lower level. The variation between replicates is also high, and so, we have at least ten replicates for each condition. Even with all these limitations, I am able to get a meaningful result from this analysis.

But the rlog transformation is not optimal for my analysis. I get a warning that more than 10% of the genes have outliers and it suggests doing vst.The vst works for without any warning, but I am still worried if this is optimal or if there is a way to do a better transformation for making heat maps and PCA.

 Also, is there a way to extract the normalized values without any transformation? DESeq used to output this, but this function is not in the DESeq2 vignette.

Thanks for your help!

SunilSukumaran

Research Associate

Monell Chemocal Senses Center

Philadelphia

USA

If you just want normalized counts, use

counts( dds, normalized=TRUE )

This hasn't changed since DESeq1. 

If you need advice on using VST or rlog, please first tell us what downstream analysis you want to perform with the transformed data.

Thanks a lot for for your reply, Simon.

I want to do a principal component analysis, calculate Pearson's correlation coefficient between pairwise comparison of all samples, and a hierarchical clustering and heat map where I use only a subset of genes that pass a  differential expression cutoff (typically fold change +/-2, padj 0.05.

Reading through some other posts, I noticed some cases where using  DESeq(dds, betaPrior=FALSE) is advised in cases where differences between the two conditions are very large. Do you advice trying this setting?

Thanks

Sunil