DESeq2 multiple levels

id0 · 05-29-2014, 02:03 PM

I loaded a set of samples into DESeq2 with one factor that has three levels. There is a single condition I called "Type" with three levels A, B, C.

This is what I did:

Code:

> ddsHTSeq <- DESeqDataSetFromHTSeqCount(sampleTable = sampleTable, directory = directory, design = ~ Type)
> dds <- DESeq(ddsHTSeq, betaPrior=FALSE)
> resultsNames(dds)
 [1] "Intercept"             "Type_B_vs_A"    "Type_C_vs_A"

I can use contrasts to run any of the pairwise comparisons. What if I want to get the differences between all three factors? Is that possible? What is the proper approach for that?

Michael Love · 05-30-2014, 08:40 AM

See the LRT section of the main vignette:

Quote:

3.6 Likelihood ratio test

One reason to use the likelihood ratio test is in order to test the null hypothesis that log2 fold changes for multiple levels of a factor ... are equal to zero.

id0 · 05-30-2014, 12:22 PM

Thanks for pointing that out. I didn't notice that before.

The example output is:

Code:

## log2 fold change: condition treated vs untreated
## LRT p-value: '~ condition' vs '~ 1'
## DataFrame with 2 rows and 6 columns
##             baseMean log2FoldChange lfcSE     stat      pvalue    padj
##             <numeric><numeric>      <numeric> <numeric> <numeric> <numeric>
## FBgn0000003   0.159  15.0447        195.974   0.791     0.374     NA
## FBgn0000008   52.226 0.0281         0.298     0.010     0.920     0.971

It says p-value is "'~ condition' vs '~ 1'", but what exactly does that mean?

Michael Love · 05-30-2014, 12:31 PM

The likelihood ratio test is a standard test, I'd recommend consulting a statistics textbook.

Or you can also start here:

http://en.wikipedia.org/wiki/Likelihood-ratio_test

~ condition is what we call the "full" model, it is also called the alternative model

~ 1 is what we call the "reduced" model, it is also called the null model

Here we are testing whether the explanatory power (in terms of the likelihood of the observed data) of including the condition information in the model, when compared to the null model (~ 1 means we fit only an intercept term, i.e., all samples belong to the same group), is more than expected by chance alone.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
DESeq2	Simon Anders	Bioinformatics	123	07-06-2015 02:45 AM
Duplication levels	ege	RNA Sequencing	4	05-23-2014 09:16 AM
DESeq2 Contrast	Ruhi	Bioinformatics	6	04-17-2014 07:12 AM
Getting started with DESeq2	ThePresident	Bioinformatics	9	02-27-2014 05:40 PM

05-29-2014, 02:03 PM	#1
id0 Senior Member Location: USA Join Date: Sep 2012 Posts: 130	DESeq2 multiple levels I loaded a set of samples into DESeq2 with one factor that has three levels. There is a single condition I called "Type" with three levels A, B, C. This is what I did: Code: > ddsHTSeq <- DESeqDataSetFromHTSeqCount(sampleTable = sampleTable, directory = directory, design = ~ Type) > dds <- DESeq(ddsHTSeq, betaPrior=FALSE) > resultsNames(dds) [1] "Intercept" "Type_B_vs_A" "Type_C_vs_A" I can use contrasts to run any of the pairwise comparisons. What if I want to get the differences between all three factors? Is that possible? What is the proper approach for that?

05-30-2014, 12:31 PM	#4
Michael Love Senior Member Location: Boston Join Date: Jul 2013 Posts: 333	The likelihood ratio test is a standard test, I'd recommend consulting a statistics textbook. Or you can also start here: http://en.wikipedia.org/wiki/Likelihood-ratio_test ~ condition is what we call the "full" model, it is also called the alternative model ~ 1 is what we call the "reduced" model, it is also called the null model Here we are testing whether the explanatory power (in terms of the likelihood of the observed data) of including the condition information in the model, when compared to the null model (~ 1 means we fit only an intercept term, i.e., all samples belong to the same group), is more than expected by chance alone.