[Simon Anders]

Simon, do you mean the estimateDispersions?

No, I mean a scatter plot of the reads.

Try, e.g.,

plot( log10( 1 + counts(cds)[1,] ), log10( 1 + counts(cds)[2,] ), pch="." )

to plot the raw, unnormalized read counts of the second sample versus the first on a log scale.

[Marianna85]

[Marianna85]

This is the script I used

CountTable=read.table("decEggs.txt", header=TRUE, row.names=1 )
head(CountTable)
decDesign = data.frame(row.names = colnames( CountTable ), condition = c( "stripped", "spawned" ), libType = c( "paired-end", "paired-end" ) )
decDesign
pairedSamples = decDesign$libType == "paired-end"
condition = decDesign$condition[ pairedSamples ]
library( "DESeq" )
cds = newCountDataSet( CountTable, condition )
cds = estimateSizeFactors( cds )
sizeFactors( cds )
head( counts( cds, normalized=TRUE ) )

and the size factors have a 100 fold difference...
what should I do???

[Marianna85]

Have you already looked at a scatter plot comparing the counts for the two samples? This should clarify what is going on.

Simon, do you mean the estimateDispersions?

[Simon Anders]

Have you already looked at a scatter plot comparing the counts for the two samples? This should clarify what is going on.

[Marianna85]

Something is going wrong somewhere, 36M and 64M reads should give normalization factors with less than a 2-fold difference. The normalization factors you listed had a 50 fold difference and suggest a much greater difference in read count.

Hi Jeremy,
in fact I was surprised to obtain such a difference...
I've not yet understood which is the mistake in the size factor calculation.

[Jeremy]

Something is going wrong somewhere, 36M and 64M reads should give normalization factors with less than a 2-fold difference. The normalization factors you listed had a 50 fold difference and suggest a much greater difference in read count.

[Shanrong]

I am using both edgeR and DESeq to analyze my own dataset. In general, the fold change reported by both are close (should be, right?).

If your orignal librares have 36 vs 64 million of total reads, your normalization factors from both methods are very weird (at least quite unusual). Please check whether you analyze your dataset properly.

The idea of normalization behind edgeR and DESeq is very similar to each other (but implementation is different). In practice, I don't see one method is superior than the other. However, it is definitely a mistake if we feed DESeq with the normalization factor from edgeR, and vice versa.

[Marianna85]

Hi Simon,
happy to read your answer

This is because edgeR gives is norm factors relative to the total read count. To get expression values on a scale comparable across sample, you have to divide the counts
- for DESeq, just by the size factod
- for edgeR, by the total read counts and by the normalization factors

Just an example to better understand.
For DESeq
gene A raw counts; 5 reads in library 1 - 70 reads in library 2
library 1:36 million reads size factor: 0.09
library 2:64 million reads size factor: 10
gene A normalized counts lib 1=5/0.09 - lib2=70/10

For edgeR
gene A raw counts; 5 reads in library 1 - 70 reads in library 2
library 1:36 million reads size factor: 0.14
library 2:64 million reads size factor: 7.2
gene A normalized counts lib 1=(5/36 million)/0.14 - lib2=(70/64million)/7.2

in this case with a very huge difference in library size, it seems better to normalize with edgeR. Isn't it?

Thanks a lot.
I really appreciate your answer.

Marianna

[Simon Anders]

With bot the normalization methods I obtain size factors very different:
-using DESeq 0,095 for one library and 10,85 for the other.
-using edgeR 0,14 and 7,2 respectively.

This is because edgeR gives is norm factors relative to the total read count. To get expression values on a scale comparable across sample, you have to divide the counts
- for DESeq, just by the size factod
- for edgeR, by the total read counts and by the normalization factors

[Marianna85]

Hi Jeremy,
yes I have a huge difference in read count between conditions: 36 vs 64 million of total reads.
I do not have other options unfortunately and I want to make a simple differential espression analysis, maybe with few differential expressed genes.
I know that, without replicates, it's difficult to make a DE analysis and I don't want to reach false conclusions. I know I have to be very conservative to say something really reliable...but how??
In your opinion, can I discard some genes, for example those with a low count reads, and make the normalization for the remaining ones?

Thanks a lot
Marianna

[Jeremy]

Hi everyone,
I'm dealing with the two normalization methods DESeq and edgeR.
I have two conditions and only one replicate per condition (I know, bad experimental design...) and I tried to normalize the raw counts.
With bot the normalization methods I obtain size factors very different:
-using DESeq 0,095 for one library and 10,85 for the other.
-using edgeR 0,14 and 7,2 respectively.

Obviously, by dividing the raw counts for the corrisponding size factor, the raw counts drammatically change, sometimes inverting the starting conditions (an upregulated gene become dowregulated).

Does it make sense?
do you think it's correct to use this normalization methods despite the weird results??

Thank you all

Looks like you have a huge difference (is that a 20-50 fold difference?) in read count between conditions, this is a problem because the normalization will significantly amplify the noise of the smaller sample making the (already unreliable without replicates) data less reliable.
But yes, you need to normalize. Is more sequencing an option?

[Marianna85]

Hi everyone,
I'm dealing with the two normalization methods DESeq and edgeR.
I have two conditions and only one replicate per condition (I know, bad experimental design...) and I tried to normalize the raw counts.
With bot the normalization methods I obtain size factors very different:
-using DESeq 0,095 for one library and 10,85 for the other.
-using edgeR 0,14 and 7,2 respectively.

Obviously, by dividing the raw counts for the corrisponding size factor, the raw counts drammatically change, sometimes inverting the starting conditions (an upregulated gene become dowregulated).

Does it make sense?
do you think it's correct to use this normalization methods despite the weird results??

Thank you all

[chadn737]

Yes, both DESeq and edgeR have functions to normalize the data. However, it's wrong to assign the size factors calculated in edgeR to DESeq, though conceptually fine at first sight. Because in DESEq, the size factor is used to 'transform' the raw reads into a 'common' ground, and you can use the normalized counts for differential analysis. But the size factor in edgeR adjusts the library size so that the gene abundence (=counts/"effective library size", and "effective library size = "library size" * "size factor") is comparable across samples.

To illustrate this point, see example below.

# data
y <- x <- rep(1,100)
y[1] <- 101  
xy <- data.frame(x=x,y=y)

#edgeR
edger <- DGEList(counts=xy)
edger <- calcNormFactors(edger)
edger$samples

#DESeq
deseq = newCountDataSet( xy, conditions=c("c1","c2") )
deseq = estimateSizeFactors( deseq )
sizeFactors( deseq )

> sizeFactors( deseq )
x y
1 1
> edger$samples
group lib.size norm.factors
x 1 100 1.4142
y 1 200 0.7071

Thank you for this. For my own work I have not done this, but in a project where I am a collaborator, the statistician in the group did use the EdgeR normalized data for input into DESeq. I know it gives very different results, and have avoided it in my own work because the DESeq size factors seemed to give more conservative results and I prefer working with fewer genes that I am very confident in than more genes of lower confidence. I'll have to bring this up on the project that I am collaborating on.

[Shanrong]

size factors in DESeq and edgeR

When you first use DESeq, you combine a table of counts and a list of conditions to create a count data set

cds <- newCountDataSet(countTable,conditions)

You can give the count data set your own size factors using

sizeFactors(cds) <- #input

If you wanted to use TMM normalized sizeFactors from EdgeR rather than those given by DESeq then you can first:

x <- calcNormFactors(as.matrix(countTable)

and then give this to the count data set:

sizeFactors(cds) <- x

Yes, both DESeq and edgeR have functions to normalize the data. However, it's wrong to assign the size factors calculated in edgeR to DESeq, though conceptually fine at first sight. Because in DESEq, the size factor is used to 'transform' the raw reads into a 'common' ground, and you can use the normalized counts for differential analysis. But the size factor in edgeR adjusts the library size so that the gene abundence (=counts/"effective library size", and "effective library size = "library size" * "size factor") is comparable across samples.

To illustrate this point, see example below.

# data
y <- x <- rep(1,100)
y[1] <- 101  
xy <- data.frame(x=x,y=y)

#edgeR
edger <- DGEList(counts=xy)
edger <- calcNormFactors(edger)
edger$samples

#DESeq
deseq = newCountDataSet( xy, conditions=c("c1","c2") )
deseq = estimateSizeFactors( deseq )
sizeFactors( deseq )

> sizeFactors( deseq )
x y
1 1
> edger$samples
group lib.size norm.factors
x 1 100 1.4142
y 1 200 0.7071