finding DE genes from VSD normalized data

[a_mt]

Hi all,

I have time series mRNA-seq data but without replicates.
I have some doubts about calling DE genes with DESeq.

I have used

varianceStabilizingTransformation

function from DESeq to normalize count data. Now can I use this vsd transformed data to calculate fold change and to call DE genes?? may be using classic LIMMA package.. Is it good practice to do so ??

I have tried

nbionTest

on raw count too.. but after vsd transforming, data look more like microarray and I was wondering is it of any harm to call DE genes/FC change on vsd transfored data.. since in original DESeq paper they have made clear that count data follows poisson distribution unlike microarray which is more like normally distributed, but after vsd transformation, data looks more like normally distributed.

Thank you.

[chadn737]

You should do replicates. And DESeq uses the negative binomial, not the poisson.

[a_mt]

Sorry, I meant to say in general count data follows poisson distribution.
But,is it ok to use vsd normalized data to detect DE genes ?? and I don't have any replicates..

[chadn737]

Does count data follow a poisson distribution? The authors of DESeq, EdgeR and others would disagree with that.

And the purpose of VSD normalized data is not for calling differential expression, but for clustering, creating heat maps, etc. In the DESeq vignette they actually describes a protocol for analyzing data without replicates, however that does not mean you should! I honestly don't know how somebody would publish results without replicates, your really can't make sense of the data without them.

[a_mt]

ok.. not arguing.. but for your reference

http://www.biomedcentral.com/1471-2105/14/91#B7

and a quote from DESeq paper :

If reads were independently sampled from a population with given, fixed fractions of genes, the read counts would follow a multinomial distribution, which can be approximated by the Poisson distribution.

quote from DEGseq paper :

Current observations suggest that typically RNA-seq experiments have low technical background noise (which could be checked using DEGseq) and the Poisson model fits data well.

And even I think no replicate does not make any sense.. but the data I am using is a published one,and just I am trying out different methods to call DEG's.

[chadn737]

ok.. not arguing.. but for your reference

http://www.biomedcentral.com/1471-2105/14/91#B7

and a quote from DESeq paper :



quote from DEGseq paper :



And even I think no replicate does not make any sense.. but the data I am using is a published one,and just I am trying out different methods to call DEG's.

I apologize if I came across a little strongly.

From the DESeq paper:

"However, it has been noted [1,8] that the assumption of Poisson distribution is too restrictive: it predicts smaller variations than what is seen in the data. Therefore, the resulting statistical test does not control type-I error (the probability of false discoveries) as advertised."

In other words, the Poisson distribution leads to false positives and is not suitable. That is why DESeq is based on a Negative Binomial, not a Poisson distribution:

"To address this so-called overdispersion problem, it has been proposed to model count data with negative binomial (NB) distributions [9], and this approach is used in the edgeR package for analysis of SAGE and RNA-Seq [8,10]."

The DESeq vignette provides protocols for analyzing data without technical replicates.

Go here: http://bioconductor.org/packages/rel.../doc/DESeq.pdf

and read section 3.3 titled "Working without any replicates." That will tell you how to do this in DESeq. The purpose of the VSD normalized data is to put everything on the same scale for clustering and other sorts of analysis, not for differential expression.