The Best way to normalize miRNA HTS data

[NicoBxl]

Hi,

I'm actually working on HTS miRNA data. I've 8 samples with about 20 M reads per file.
After adapter trimming in each sample, the number of reads are different in each sample (each sample has a different quality (preparation and sequencing)).

Example :

file 1 : 200 000 reads
file 2 : 1 000 000 reads
...

Now I want to analyze the differential expression. The problem for doing that is the normalization step . What is the best way to normalize the data for comparing multiple samples ?

Thanks a lot,

N.

[NicoBxl]

After searching I found some methods :

- RPM ( Read Per Million )
- RPKM (Reads per kilobase per million mapped) : Mortazi et al, Nat. Methods, 2008
- Trimmed mean of M-values : Robinson, Oshlack, Genome Biology 2010
- upper-quantile : http://www.ncbi.nlm.nih.gov/pubmed/20167110

Which method is the best for HTS miRNA data ?

[Simon Anders]

Dividing by the number of sequenced or mapped reads is a bad idea, for the reasons explained by Robinson and Oshlack.

I'd advise against quantile normalization (as suggested by Bullard et al., the fourth in your list) . RNA-Seq is known to be linear, and quantile normalization will only distort this. This leaves you with TMM (Robinson and Oshlack) or with the method that we implemented in DESeq (preprint here), which is similar in spirit to TMM but uses a bit different math.

This all is said assuming that you want to make comparisons between samples, i.e., see whether a given gene's expression depends on your experimental conditions. If you want to compare different genes within the same sample, you need a very different approach. In that case, look at cufflinks (Trapnell et al, 2010).

Simon

[Simon Anders]

This all is said assuming that you want to make comparisons between samples, i.e., see whether a given gene's expression depends on your experimental conditions. If you want to compare different genes within the same sample, you need a very different approach. In that case, look at cufflinks (Trapnell et al, 2010).

Sorry, this paragraph was nonsense; I forgot that you aredealing with miRNA. The whole point of cufflinks (an its FPKM measure) is to deal with splicing variants and differing transcript length. This is not an issue if the read length exceeds the transcript length as is the case with miRNA.

Simon

[NicoBxl]

ok thanks Simon,

I'll try DESeq

Nicolas

[NicoBxl]

Is it possible wit DESeq to get the normalized count matrix ? to draw a heatmap with the normalized data per example

[Simon Anders]

Is it possible wit DESeq to get the normalized count matrix ?

Yes, just divide the jounts by the size factors:

t( t(counts(cds)) / sizeFactors(cds) )

(All the 't are to make sure that R divides by column, not by row.)

to draw a heatmap with the normalized data per example

You might want to take the log or use DESeq's variance-stabilizng transformation for such a heatmap. (Careful with the latter; I've just found a bug in 'getvarianceStabilizedData'. It's fixed in the devel version (1.1.11) of DESeq but not yet in the release branch.)

Simon

[NicoBxl]

ok thanks,

I'll try that.

For the heatmap, it's the log(normalized counts) that you talking about ?

For the variance, I'll download the dev version

[NicoBxl]

Yes, just divide the jounts by the size factors:
You might want to take the log or use DESeq's variance-stabilizng transformation for such a heatmap.

I don't understand how to get the log ? Can you explain me that Simon?

Thanks a lot

N.

[cascoamarillo]

Hi guys,

I'm also trying to normalized the diff expression in my small RNAs. I'd like to try DESeq in R; but after looking of its info:

http://www-huber.embl.de/users/anders/DESeq/

I can't understand how could I do this analysis. Is there any DESeq guide for Dummies? does it take any of my bam files???

Sorry to bother you guys with these silly questions.

[Simon Anders]

I'm also trying to normalized the diff expression in my small RNAs. I'd like to try DESeq in R; but after looking of its info:
http://www-huber.embl.de/users/anders/DESeq/

Have you even tried to read the manual ("vignette")? I ask because the URL you have put does not contain much useful information in addition to the link to the manual, but it does not sound as if you followed it.

[cascoamarillo]

Have you even tried to read the manual ("vignette")? I ask because the URL you have put does not contain much useful information in addition to the link to the manual, but it does not sound as if you followed it.

Thanks for the reply.

Yes, I've been looking the manual. Maybe the problem is my little experience with R. I made same plots and HTS data analysis in R, but not an expert on it. Sorry to express my frustration in the post. I'll try to look at the manual further and see what happen.

[cascoamarillo]

Ok, after a further reading of the manual ("vignette") I've some questions. First of all, I don't work with Drosophila, so I do not need the library pasilla, right?. What I did was to take my sam and gff file and create the count table of my treatened/untreatened conditions (HTSeq). But how can I load this table into DESeq? Do I need my own library "pasilla" ?

Thanks

[Simon Anders]

No, the 'pasilla' library is example data. It comes with another vignette which, among other things, tells you how to get the data into R.

On the long run, it may be well worth your while to read some short introductory tutorial to R. Reading in data tables and writing them out again is such a common procedure that hardly any documentation for a specific R package will explain you how this is done. When I write things like the DESeq vignette, I assume that the reader has already familiarized himself with these basics.

(I know that a lot of people here hope for single-click push-button solutions. You will get them --- in ten years or so, when analysis methods are no longer new and constantly changing, but well-matured textbook knowledge with fixed consensus recipes that can be followed blindly without a need to understand them. Until then, you will need some basic statistics and bioinfomatics knowledge to make informed choices about the many different possibilities to analyse your data.)

[cascoamarillo]

Ok, there's another DESeq manual. I guess this is the one you are talking about:

http://bioconductor.org/packages/2.8/bioc/vignettes/DESeq/inst/doc/DESeq.pdf

The other is:

http://www.bioconductor.org/packages/devel/bioc/vignettes/DESeq/inst/doc/DESeq.pdf

Looks like the same but not; sorry for the confusion. "Analysing RNA-Seq data with the “DESeq” package" is the one that I need to start working with DESeq. Thanks a lot for this package and the complete documentation that you've provided.