How to Normalize NGS data? Tags per million?

[xhuister]

Hi all,

I have three libraries from 3 experiments. Our sequences are tags with a poly(A) tail, like the polyA-tail sequences in mRNA-seq but using different protocol. Most of the tags are located in 3'-UTR regions.

I think I need to normalize the data before comparing different libraries.

There is a normalization method called TPM:
TPM (tags per million) according to the total tag count in each library
as follows: TPMj i=Cji*10^6/Tj, where TPMj i is the TPM for SAGE tag i in SAGE library j, Cj i is the count of SAGE tag i in library j, and Tj is the total SAGE tag count in SAGE library j.

It seems that this method is just to normalize the sequence number by the total sequence count in each library.

But, if I use TPM, should I use the raw sequence count or only the count of the mapped sequence? Or any other normalization method?

Thanks a lot.

[krobison]

I would think you would want to use mapped tags as the denominator; otherwise poorly sequenced/prepared libraries will artificially have lower expression values

[Wolfgang Huber]

Form follows function

Hi Xhuister

when asking 'how' to do normalisation it is a good idea to first ask 'what for'. In RNA-Seq, a typical reason is to avoid spurious differential expression calls just because of differential library coverage. Another consideration is that you want to keep track of the actual counts when assessing statistical confidence in differential expression calls, since the counting noise is relatively more important when the numbers are small, even for the same fold-change.

A paper by M. Robinson and A. Oshlack discusses especially the first aspect in quite some detail: http://genomebiology.com/2010/11/3/R25

A paper by S. Anders and myself combines a very similar normalisation method with the error modeling needed for confidence computations: http://precedings.nature.com/documents/4282/version/2

Best wishes
Wolfgang Huber

[xhuister]

Thank you Wolfgang and Krobison,

I'm reading the DEG paper now. But I'm not sure whether it is suitable for my case.

In my case, for each gene, there are only some (normally <5) locations with reads and most of the locations are in 3'-UTR, not like the case that the reads are distributed along the transcripts.

Do you think it's OK to use the normalization method using DESeq or just use 'Tag per million' to normalize by the total count in each library? Thank you!

[Wolfgang Huber]

Dear Xhuister

as long as you have reason to believe that your counts are roughly proportional to the true target gene abundance (with unknown proportionality factors that depend on sample, lane and gene), these normalisations are in principle suitable. (And if that were not the case, then I am not sure what you want to normalise.)

Best wishes
Wolfgang

[xhuister]

Thank you Wolfgang. Maybe I'll have a try both TPM and DESeq.