DESeq without replicates

[austinpa]

Hello,

I am unable to get method=blind to work to calculate variance without replicates. Or actually any "method" argument from the vignette. Is this now obsolete? Are the only variance calculations supported now "pooled" or with replicates? Has anyone else had trouble with this recently?

Thanks,
Austin

[RockChalkJayhawk]

Hello,

I am unable to get method=blind to work to calculate variance without replicates. Or actually any "method" argument from the vignette. Is this now obsolete? Are the only variance calculations supported now "pooled" or with replicates? Has anyone else had trouble with this recently?

Thanks,
Austin

try

 method="blind"

[Simon Anders]

... and make sure you use a current version of DESeq. I've added the 'method' argument only a few months ago.

[RockChalkJayhawk]

... and make sure you use a current version of DESeq. I've added the 'method' argument only a few months ago.

Do you also have some more documentation on the 'method' argument? I can't seem to find it.

[Simon Anders]

Do you also have some more documentation on the 'method' argument? I can't seem to find it.

Start R, load DESeq, and type "?estimateVarianceFunctions". If you don't see anything there about 'metho', you have an old DESeq version.

Simon

[austinpa]

DESeq without replicates

Great. Thanks Simon. I just needed the newer version of DESeq and R.

Austin

... and make sure you use a current version of DESeq. I've added the 'method' argument only a few months ago.

[taoxiang180]

Hi, everyone.
I am a nevice to R and DESeq. Working without any replicates, I went through each step of differential gene expression analysis from the DESeq manual, But the results really puzzled me: some gene have a pval<0.05, but have a padj of 1, how the padj was calculated? and how to set the padj threshold？Any pointers will be appreciated.Thanks!

[taoxiang180]

Hi, everyone.
I am a novice to R and DESeq. Working without any replicates, I went through each step of differential gene expression analysis from the DESeq manual, But the results really puzzled me: some gene have a pval<0.05, but have a padj of 1, how the padj was calculated? and how to set the padj threshold？Any pointers will be appreciated.Thanks!

[RockChalkJayhawk]

Hi, everyone.
I am a novice to R and DESeq. Working without any replicates, I went through each step of differential gene expression analysis from the DESeq manual, But the results really puzzled me: some gene have a pval<0.05, but have a padj of 1, how the padj was calculated? and how to set the padj threshold？Any pointers will be appreciated.Thanks!

The padj corrects for multiple testing to limit your false discovery rate. Using this value, you can essentially interpret P<0.05 to mean that you are 95% sure there is a difference in expression.

[Simon Anders]

The padj corrects for multiple testing to limit your false discovery rate. Using this value, you can essentially interpret P<0.05 to mean that you are 95% sure there is a difference in expression.

Not quite. Especially, your phrasing "95% sure" is to imprecise to capture the difference between an unadjusted p value and one adjusted for FDR control.

I've just googled a bit, looking for a good primer to explain this concept (which is of vital importance: don't even think about analysing genomics data if you don't know what the multiple hypothesis testing problem is), but most of the stuff is too technical for a reader without statistics training. This paper here might explain it reasonably well, but is a bit lengthy:

Pounds SB. Estimation and control of multiple testing error rates for microarray studies.
Briefings in Bioinformatics. 2006;7(1):25-36.

http://bib.oxfordjournals.org/cgi/content/abstract/7/1/25

(DESeq uses Benjamini and Hochberg's method. See the article to learn what this means.)

Simon

[RockChalkJayhawk]

Not quite. Especially, your phrasing "95% sure" is to imprecise to capture the difference between an unadjusted p value and one adjusted for FDR control.

I've just googled a bit, looking for a good primer to explain this concept (which is of vital importance: don't even think about analysing genomics data if you don't know what the multiple hypothesis testing problem is), but most of the stuff is too technical for a reader without statistics training. This paper here might explain it reasonably well, but is a bit lengthy:

Pounds SB. Estimation and control of multiple testing error rates for microarray studies.
Briefings in Bioinformatics. 2006;7(1):25-36.

http://bib.oxfordjournals.org/cgi/content/abstract/7/1/25

(DESeq uses Benjamini and Hochberg's method. See the article to learn what this means.)

Simon

Thanks for pointing that out. That's what I get for not finishing my coffee before I type.

Perhaps a better way to say it is that the padj changes the pvalues so that only 5% are likely false positives - it has nothing to do with the specific gene being differentially expressed.

[Simon Anders]

Exactly.

So for those to lazy to read the paper:

If you have 10,000 genes and you use a threshold of 0.05 on the raw p values, you should get around 500 false positives (5% of 10,000). So if there are 1000 genes with p<0.05, about half of them might be false positives.

If you use a threshold of 0.05 on the adjusted p values, you will find fewer genes, let's say, 100, but now, you know only 5% of these are false positives (here: 5 of 100 genes).

Simon

[Jeremy]

The main problem is no replicates, this means the variance is calculated using both your samples pooled so a large difference will have a large variance and thus less statistical significance. You may want to ignore the p-values completely and just look at highest and lowest fold change, because you have no replicates you will need to confirm pretty much everything with (replicate) qPCR.

[taoxiang180]

Thanks for so much recommendation！
When work without any replicates and have a padj=1, what can I do next using DESeq?
(Obviously, we can not make a conclution that there are no differential expression between two groups)
According to the Yoav Benjamini's method(2001), once the P-value are
available from any statistical software, the extra FDR calculation can be done easily within a spreadsheet software such as excel using the built-in functions, can I calculate FDR in this way?

[Simon Anders]

As Jeremy said, adjusted p values are of little use in your situation. Doing experiments without replicates is simply a bad idea.

Just start from the largest fold change (or maybe from the smallest unadjusted p value) and keep doing qPCR (with at least, say, five, biological replicates!) for all genes until you run out of patience.

And you don't have to use Excel, you can use R's 'p.adjust' function for the multiple testing calculation, For your convenience, DESeq runs 'p.adjust' with 'method="BH"' (Benjamini-Hochberg) on the 'pval' column of the result and reports this in the 'padj' column.

Alternatively, you may want to give this, quite new, idea a try:

Wu Z, Jenkins BD, Rynearson TA, et al. Empirical Bayes Analysis of Sequencing-based Transcriptional Profiling without Replicates.
BMC Bioinformatics. 2010;11(1):564. http://www.biomedcentral.com/1471-2105/11/564