Question related to analysis using DESeq2

[kjaja]

Hi All,

I have expression data for a mutant where a drug is injected over time with increasing dosage (no replicates and no drug was injected at time 0). I would like to identify genes that change in expression significantly over time. I used the following design in DEseq2 where time is assumed to be a continuous variable.

ddsp <- DESeqDataSet(dds, ~ time)
ddsp <- DESeq(ddsp, test="LRT", reduced = ~ 1)
resp <- results(ddsp)

is this the correct design? since I have only one mutant, are there other possible ways to answer the same question?

[Michael Love]

If time is a numeric vector:

class(dds$time)

then this will find genes which have linear relationship between log counts and time.

(note to other people reading the thread: if you have replicates, it's preferable to include time as a factor, such that each time point gets its own parameter in the model)

If you have more than 3 time points, you can include a quadratic term, which will also find genes which have up-down or down-up patterns in addition to linear relationships. That would look like a design of ~ time + I(time^2) and reduced design ~ 1.

[kjaja]

Thanks Michael ,

To follow up on my original post, I am trying to compare the results between DESeq and DESeq 2 and found the output was very different (correlation=0.59), is that expected? I have read other posts and learned that we expect more genes to be differentially expressed using DESeq2 compared to DESeq which is what I got. The code for DESeq follows


### using DESeq ##############
cds = newCountDataSet(data, expdesign)
cds <- estimateSizeFactors( cds )
sizeFactors( cds )
cds = estimateDispersions( cds, method="blind", sharingMode="fit-only" )
fit1 <- fitNbinomGLMs( cds, count ~ time )
fit0 <- fitNbinomGLMs( cds, count ~ 1 )
res<-nbinomGLMTest( fit1, fit0 )

[kjaja]

question related to analysis using DESeq2

Hi,

when using the following design to test for differentila expression

ddsTC <- DESeqDataSet(dds, ~ time)
ddsTC <- DESeq(ddsTC, test="LRT", reduced = ~ 1)
resTC <- results(ddsTC)

How to interprest the log2foldchange in the results file?
I find most of the genes with significant pvalues have a very low log2fold change.

thanks

If time is a numeric vector:

class(dds$time)

then this will find genes which have linear relationship between log counts and time.

(note to other people reading the thread: if you have replicates, it's preferable to include time as a factor, such that each time point gets its own parameter in the model)

If you have more than 3 time points, you can include a quadratic term, which will also find genes which have up-down or down-up patterns in addition to linear relationships. That would look like a design of ~ time + I(time^2) and reduced design ~ 1.

[Michael Love]

Read the section on the likelihood ratio test in the help pages:

?results

[kjaja]

Thanks Michael for your help
I have one more question , do I need to perform a between-sample normalization before using the following design ?

ddsTC <- DESeqDataSet(dds, ~ time)
ddsTC <- DESeq(ddsTC, test="LRT", reduced = ~ 1)
resTC <- results(ddsTC)

thanks

[kjaja]

Any help will greatly be appreciated.

[fanli]

the 'DESeq' function does normalization internally