[Marianna85]

sorry

http://dl.dropbox.com/u/33322766/plot.JPG

[Marianna85]

sorry again...i think it's not ok

[Marianna85]

https://dl.dropbox.com/u/33322766/plot.jpg

[Simon Anders]

Wow, this is quite a mess. Huge amounts of points with 0 reads in one and more than 10,000 reads in the other library. What have you sequenced there? If this is ordinary RNA-Seq data, things went quite wrong.

[Marianna85]

this is a RNAseq experiment in mollusc eggs. Raw reads after orthologs clustering. Yes, there are some points with a huge different among the two libraries. Is this the reason why I obtained so different normalization factors?

[Marianna85]

But...the two samples correspond each to a different stage...Maybe the expression profile between the two conditions is huge...and in this case, which is the best solution in order to normalize the raw count?

[Simon Anders]

No matter how you normalize, all genes will show huge differences. And you will have no way to figure out whether you might have seen as huge differences has you sequenced two eggs from the same developmental stage (this is well possible: maybe mollusc eggs differ a lot from each other, or maybe your wet-lab protocol is unstable). This is, of course, just another example while doing an experiment without at least duplicates is simply bad science and a waste of time and money (and by now, no decent journal will accept such studies any more, I hope), but I am repeating, what I already said often in previous threads.

Sorry for the pessimism, but quite frankly, I doubt that it's worth putting much more effort into this analysis. What exactly had you hoped to find?

[Marianna85]

No matter how you normalize, all genes will show huge differences. And you will have no way to figure out whether you might have seen as huge differences has you sequenced two eggs from the same developmental stage (this is well possible: maybe mollusc eggs differ a lot from each other, or maybe your wet-lab protocol is unstable). This is, of course, just another example while doing an experiment without at least duplicates is simply bad science and a waste of time and money (and by now, no decent journal will accept such studies any more, I hope), but I am repeating, what I already said often in previous threads.

Sorry for the pessimism, but quite frankly, I doubt that it's worth putting much more effort into this analysis. What exactly had you hoped to find?

The experiment is mainly a transcriptome characterization...this is the reason why we didn't plan any replicates.In order to study the whole transcriptome, in my opinion, a simple normalization intra-library is enough.
But I also hoped to find few genes differentially expressed, without a p-values and without statistical power, just to find transcripts (maybe 5-10) which show a huge difference between the two conditions. Just a first step to be validated, no more. To do this I need an inter-library normalization.
Considering the data, do you think that this objective is too ambitious?
Marianna

[puggie]

I work on whole-transcriptome also. I believe there is a lot of alternatives for normalizing this kind of data. I have tried edgeR but it did not produce meaningfull normalization by human eyes. What I do instead is based on FPKM/RPKM counts e.g. via Cufflinks. E.g. select 5-10.000 genes that are expressed in all samples and then normalize by these. If samples are of different origin this approach may prove insufficient.

[Marianna85]

Hi Puggie,
thanks for your reply.

E.g. select 5-10.000 genes that are expressed in all samples and then normalize by these

what do you mean normalize by 5-10.000 gene, how? you mean RPKM normalization?

If samples are of different origin this approach may prove insufficient.

the two samples I have come from the same tissue but different maturation stage. But several genes seem to be extremely DE (0 vs 100.000 reads!)

Anyway RPKM normalization makes sense only if you don't need to compare different samples.

Marianna