[IsBeth]

Estimate Dispersions with DESeq2

Hello! =)

I'm a beginner (Blutiger Anfänger) at bioinformatics and so I've got a beginners question, sorry! Nowadays I'm trying to run DESeq2 with a dataset without replicates. It's a dataset with two columns (2 conditions, "control" and "infected").
Which method do you recommend me to estimate the dispersion of my data? ( "pooled", "pooled-CR", "per-condition", "blind"?). I was reading about these methods but it's difficult for me to say which one I should use.

If I just use the command "estimateDispersions" without specifying the method to make use of, which one is applied?

What should I be careful with?


Thanks in advance!

[Michael Love]

fpesce,

sorry I can't replicate any bug with mixed numeric and categorical variables for >100 samples. If you can email me a small reproducible example (maybe subsetting the rows of dds) I can look into it. My email is listed in help(package="DESeq2")

But I would recommend turning age into a categorical variable. You can split it into biologically meaningful groups: <18 y/o, 19-35, etc. Then the average effect of each group is controlled for, regardless of the trend over years. This often makes more sense rather than assuming a constant log2 fold change per year.

You can turn numerics into factors using cut():

> age <- 40:50
> cut(age, breaks=c(0,30,45,60,200))
[1] (30,45] (30,45] (30,45] (30,45] (30,45] (30,45] (45,60] (45,60] (45,60]
[10] (45,60] (45,60]
Levels: (0,30] (30,45] (45,60] (60,200]

[Michael Love]

IsBeth,

There is information about the methods used by estimateDispersions in the ?estimateDispersions man page.

In DESeq2, the Cox-Reid estimator (a penalized maximum likelihood estimator) is the only method provided, as it covers all experimental designs.

it is also discussed in this man page what happens in the case of a 1 vs 1 comparison.

[fpesce]

fpesce,

sorry I can't replicate any bug with mixed numeric and categorical variables for >100 samples. If you can email me a small reproducible example (maybe subsetting the rows of dds) I can look into it. My email is listed in help(package="DESeq2")

But I would recommend turning age into a categorical variable. You can split it into biologically meaningful groups: <18 y/o, 19-35, etc. Then the average effect of each group is controlled for, regardless of the trend over years. This often makes more sense rather than assuming a constant log2 fold change per year.

You can turn numerics into factors using cut():

> age <- 40:50
> cut(age, breaks=c(0,30,45,60,200))
[1] (30,45] (30,45] (30,45] (30,45] (30,45] (30,45] (45,60] (45,60] (45,60]
[10] (45,60] (45,60]
Levels: (0,30] (30,45] (45,60] (60,200]

Thank you so much Mike, this completely makes sense and actually solved this issue.

[IsBeth]

Thanks Mike for your answer =)

I should have said before that I have read already the brief explanations about the different methods (sorry), yet my problem is that I have not enough statistic skills to understand why I'm using a certain dispersion method and not another in DEseq (not DEseq2). Is there a way to get specific math and statistic skills to understand R properly?

[mbahin]

Hi all,

I am trying to use DESeq2 just after HTSeq-count.

I am following the tuto in the DESeq2 vignette but when I try to do the:

res <- results(dds)

Although I think I just followed the tuto properly, I get this error message

Error in is.numeric(cooksCutoff) : 'cooksCutoff' is missing

I can't find any answer about this error...

Cheers

[mbahin]

Hi all,

I just found the answser, I was trying to use DESeq2 with only two count files (waiting for the other one to come).
Now that I am using it with 6 files it's ok. It looks like it was the issue.

Cheers

[Michael Love]

hi,

Sorry, this was a simple bug for when number of samples equaled number of coefficients, which also came up on the Bioconductor mailing list. I pushed a fix to the release branch.

http://comments.gmane.org/gmane.science.biology.informatics.conductor/51865

Mike

[kmcarr]

Hello,

In DESeq, it was possible to view/export the baseMean values for each condition (baseMeanA and baseMeanB) in addition to the overall baseMean. In DESeq2, I have only been able to view/export the overall baseMean. Is it possible to acquire the baseMean values for each condition as well? Thanks!

I'm bumping up KHubbard's question because I have the same one and it seems it didn't get answered the first time around. I have looked at the structure of the DESeqDataSet object (best as an R neophyte can) and don't see them in there. Is there any way to get these values in DESeq2?

Thanks.

[dpryan]

I'm bumping up KHubbard's question because I have the same one and it seems it didn't get answered the first time around. I have looked at the structure of the DESeqDataSet object (best as an R neophyte can) and don't see them in there. Is there any way to get these values in DESeq2?

Thanks.

Wasn't Michael's reply (the next post) in regard to that?

[kmcarr]

Wasn't Michael's reply (the next post) in regard to that?

In the words of Homer Simpson, "DOH!!"

Thank Devon, and apologies to everyone.

[IsBeth]

Hello,

I'm using DESeq2 and so far it seems to work well. But why are the log fold changes so low? They have values between 0 and 1, even when the genes are differentially expressed.

[Simon Anders]

Maybe there is not much happening in your data?

Can you post an MA plot?

Also, make sure you understand the conspt of shrunken fold changes. (See our paper for details.)


Edit: Values between 0 and 1? No negative values? Sounds like you are looking at the p values, not the log fold changes.

[IsBeth]

In my MAplot are differential expressed genes (in red). Yes, there are also negative values. I think I don't understand the concept of shrunken fold changes, still a newbie

It just seemed strange to me as I compared the lfc values after using edgeR and DESeq2

[Simon Anders]

If you don't show us the MA plot, we can't tell you whether it looks right. That's why I suggested you post it. Also tell us about the biology of your experiment.

EdgeR's log fold-changes are not shrunken, DESeq2's are. They are a new feature in DESeq2, and explained in our paper.

Hence, sorry if the following question sounds a bit direct but: By "I don't understand the concept of shrunken fold changes", do you mean that you have read our preprint but found the explanation unclear (in which case we should improve it) or that you have not yet looked at our paper (in which case we would be disappointed that nobody reads what we wrote)? ;-) You don't need to read the whole thing; just the section on shrunken fold changes.